Smart Communications, Intelligent Algorithms and Interactive Methods: Proceedings of 4th International Conference on Wireless Communications and ... Innovation, Systems and Technologies, 257) 9811651639, 9789811651632
This book gathers selected papers presented at the 4th International Conference on Wireless Communications and Applicati
120 59 8MB
English Pages 336 [313] Year 2022
Preface
Contents
About the Editors
1 Algorithm and Application of Artificial Intelligence in Wireless Sensor Network Coverage
1.1 Wireless Sensor Network Application and Coverage
1.2 Overview of Particle Swarm Optimization Based on Artificial Potential Field
1.3 Use of Intelligent Algorithms in WSN Coverage
1.4 Conclusion
References
2 Design of Cooperative Driving System Based on VANET
2.1 Introduction
2.2 Overall System Model
2.3 Working Principle of the Collaborative Driving System
2.3.1 Vehicle Lane Change
2.3.2 Vehicle Following
2.4 System Simulation
2.5 Conclusion
References
3 A Study of Intelligent Power Information Collection System Based on Distributed Architecture
3.1 Introduction
3.2 The System Architecture
3.2.1 The System Configuration
3.2.2 The Master Station Layer
3.2.3 The Communication Layer
3.2.4 Intelligent Terminal
3.3 Implementation of Intelligent Power Consumption Information Acquisition System
3.3.1 The Physical Architecture
3.3.2 Internal and External Contact
3.4 Conclusions
References
4 A Multiuser Communication Algorithm for UAV Bee Combat
4.1 Introduction
4.2 Chirp Signal Against Doppler Frequency Shift
4.2.1 Chirp Signal Impulse Compression Characteristic
4.2.2 Analysis of Chirp Signal Anti-Doppler Performance
4.3 Division-Time Multiuser Chirp Modulation Algorithm
4.4 Simulation Results
4.4.1 Gaussian White Noise Channel
4.4.2 Doppler Channel
4.5 Conclusions
References
5 Review on Dec-POMDP Model for MARL Algorithms
5.1 Introduction
5.2 Multi-agent Reinforcement Learning
5.2.1 Background of MARL
5.2.2 Modeling for MARL Problem
5.3 Advances in MARL Algorithms
5.3.1 MARL Algorithm of Communication-Based Learning
5.3.2 MARL Algorithm of Collaborative-based Learning
5.4 Challenges and Research Directions for MARL Algorithms
5.5 Conclusion
References
6 Research on AGV Correction Control Method Using QR Code Positioning
6.1 Introduction
6.2 AGV Correction Model Establishment
6.3 AGV Pose Correction System
6.3.1 QR Code Deviation Parameter Calculation
6.3.2 Research on Correction Algorithm of AGV Motion Track
6.4 AGV Operation Experiment
6.5 Conclusion
References
7 Research on Multi-data Center Collaboration Technology for Multi-station Fusion
7.1 Introduction
7.2 Resource Management Architecture of Distributed Data Center and Core Data Center Platform
7.3 Multi-level Allocation and Recycling Technology for Resources in Multiple Data Centers
7.4 Data Synchronization Technology for Collaborative Business Between Distributed Data Center and Core Data Center Platform
7.5 Conclusion
References
8 A Novel Method for Online Control in Equipment Parameters of Blow Molding Process
8.1 Introduction
8.2 Product Quality Prediction Mode
8.2.1 Orthogonal Design
8.2.2 Random Forest Prediction Model
8.2.3 Process Parameter Correlation Analysis
8.3 Process Parameter Collaborative Control
8.3.1 Process Parameters Collaborative Control Specific Process
8.3.2 Experimental Verification
8.4 Conclusion
References
9 Research on the Optimum Ratio of Intelligent Rail Guided Vehicle and Computer Numerical Control
9.1 Introduction
9.2 Problem Analysis
9.3 Model Building
9.3.1 Determination of RGV Movement Time
9.3.2 Estimation of CNC Upper Limit
9.3.3 A Preliminary Division of Processing Time T for a Single Material
9.3.4 The Optimal Path Planning Model
9.3.5 Determination of Optimum Ratio of CNC
9.4 Solution and Result
9.4.1 Solution of the Optimal Path Planning Model
9.4.2 The Solution of Optimal CNC Number Ratio
9.5 Conclusions
References
10 An Apple Grading Method Based on Improved VGG16 Network
10.1 Introduction
10.2 Material Preparation
10.2.1 Apple Grading Standards
10.2.2 Data Acquisition
10.2.3 VGGNet
10.3 Improvement of VGG16 Architecture
10.4 Experiment of Apple Grading
10.4.1 Hardware Platform
10.4.2 Experiment Procedure
10.4.3 Training Parameters
10.4.4 Experimental Results
10.4.5 Apple Grading Based on Machine Learning Algorithms
10.5 Conclusions
References
11 Comprehensive Application of Big Data in Lean Management of Distribution Network Operation
11.1 Introduction
11.1.1 Topological Relationship Verification of Distribution Network Station Area is Time-Consuming and Laborious
11.1.2 The Operation and Maintenance Management of Existing Distribution Station Area not Enough Lean
11.1.3 The Technical Reconstruction Projects in the Existing Distribution Station Area Are not Accurate Enough
11.1.4 The Requirement of IT Capability for Mass Data Query
11.2 Data Mining and Research of Distribution Network
11.2.1 Deep Mining the Big Data of Operation and Power Distribution System to Improve the Load Forecasting Ability of Different Industries and Regions
11.2.2 Deep Mining the Load Data to Improve Access Ability of Distribution Network
11.3 Lean Management Platform of Distribution Network
11.3.1 Three-Phase Imbalance Analysis of Distribution Network
11.3.2 Voltage Quality Analysis of Distribution Network
11.3.3 Load State Analysis of Distribution Network
11.3.4 Low-Power Factor Analysis of Distribution Network
11.3.5 Reactive Power Operation Evaluation of Distribution Network
11.4 State Assessment of Distribution Network Equipment
11.4.1 State Assessment of Distribution Area—Contact State Assessment
11.4.2 Operational Status Assessment of Area-Overload Fault Risk Assessment
11.5 Conclusion
11.5.1 Enhance “Lean Management” for Operation and Maintenance of Distribution Area
11.5.2 Realize Modified “Precise Investment” of Distribution Area
References
12 Research on the Construction of Intelligent Vehicle Verification System for Road Transportation
12.1 Introduction
12.1.1 Industry Background
12.1.2 Situation Analysis
12.2 System Workflow
12.3 System Construction
12.3.1 Overall Functional Structure
12.3.2 Realization of Specific Functions
12.3.3 Function Optimization Scheme
12.4 Conclusion
References
13 Quantitative Risk Analysis of Dangerous Goods Road Transport Vehicles and Networks
13.1 Introduction
13.2 Risk Analysis Model
13.3 Conclusion
References
14 Virtual Reality Technology Facilitates Customized Clothing Design in C2M Business Model
14.1 Introduction
14.2 The Development of VR in the Field of Clothing
14.3 Virtual Reality Technology Facilitates C2M “Manufacturing End” Custom Clothing Design
14.3.1 More Vivid Model Display
14.3.2 More Intuitive and Stereoscopic Appearance
14.3.3 Realize the Transformation from 2 to 3D in the Structure Design
14.3.4 Facilitate Clothing Functional Research and Development
14.4 VR Facilitates C2M “Client End” Custom Redesign
14.5 VR Makes Shopping Experience More Interesting Under C2M Mode
14.6 Closing Remarks
References
15 Research on Identification Model of Special Transformer Stealing Electricity Based on Vector Similarity Matching Algorithm
15.1 Introduction
15.2 Analysis of Power Stealing of Special Transformer
15.2.1 Summary of Power Stealing Methods for Special Transformer
15.2.2 Summary of Power Consumption Characteristic Signals of Special Transformer Users
15.3 Current Situation of Electricity Stealing Samples of Special Transformer Users
15.4 Power Stealing Identification of Special Transformer Based on Vector Similarity Matching Algorithm
15.4.1 Vector Similarity Comparison Method
15.4.2 Identification of Power Theft by Special Transformer Users
15.4.3 Algorithmic Process Analysis
15.4.4 Application Results of Algorithmic Model
15.5 Conclusion
References
16 International Checkers Evaluation Algorithm Based on Evolutionary Neural Network
16.1 Introduction
16.2 Design Idea of International Checkers Evaluation Algorithm Based on Evolutionary Neural Network
16.3 Implementation of International Checkers Evaluation Algorithm Based on Evolutionary Deep Learning
16.3.1 Deep Neural Network Pre-model Design and Implementation
16.3.2 Design and Implementation of Evolutionary Neural Network Mode
16.4 Experimental Results and Analysis
16.5 Conclusion and Prospect
References
17 Research on Intelligent Health Diagnosis of Tunnel Structure System Based on Wavelet Neural Network
17.1 Introduction
17.2 Experimental Process
17.2.1 Test Equipment and Basic Principles
17.2.2 Specimen Production and Testing
17.3 The Establishment of Wavelet Neural Network Structure
17.4 Feature Vector Extraction of Wavelet Neural Network
17.5 Training of Wavelet Neural Network
17.6 Conclusions
References
18 Data on Cloud Practice Based on Alibaba Cloud IOT Platform
18.1 Introduction
18.2 Types of IOT Platforms
18.3 Basic Functions of Alibaba Cloud Internet of Things Platform
18.4 Basic Operation Steps of Alicloud Platform
18.4.1 Register as an Alicloud User
18.4.2 Add Products and Devices
18.5 MQTT Software Connectivity Test Steps
18.6 Conclusion
References
19 Remote Intelligent Guidance System Based on Augmented Reality
19.1 Introduction
19.2 System Design
19.2.1 System Framework
19.2.2 Hardware and Platform Selection
19.3 Software Overall Design
19.3.1 Software Design for Astronaut Part
19.3.2 Software Design for Remote Expert Part
19.3.3 Design and Simulation of Space-Earth Link
19.4 Results and Discussion
19.5 Conclusions
References
20 Research on Air Engine Air Circuit Fault Based on Convolutional Neural Network Feature Extraction
20.1 Introduction
20.2 CNN and SVM-Based Aeroengine Fault Diagnosis Method
20.2.1 Fault Diagnosis Process and Implementation Steps
20.2.2 Structural Design of Convolutional Neural Network
20.3 Verification of Aeroengine Gas Path Fault Diagnosis Method Based on CNN and SVM
20.4 Conclusion
References
21 Improved Adaptive Particle Swarm Optimization Algorithm with a Two-Way Learning Method
21.1 Introduction
21.2 Double Adaptive Optimization Strategy
21.2.1 Adaptive Inertial Weight
21.2.2 Adaptive Learning Factor
21.3 Basic Beetle Antenna Search Algorithm
21.4 Two-Way Learning Strategy
21.5 Algorithm Steps
21.6 Simulation Experiment and Result Analysis
21.6.1 Test Function
21.6.2 Test Experiment and Analysis
21.7 Conclusion
References
22 Flower Pollination Algorithm Based on Beetle Antennae Search Method
22.1 Introduction
22.2 Basic Principles
22.2.1 Flower Pollination Algorithm (FPA)
22.2.2 Beetle Antennae Search Algorithm (BAS)
22.2.3 Population Initialization Method of Uniform Sampling Using Toeplitz Matrix
22.2.4 Flower Pollination Algorithm Based on Beetle Antennae Search Algorithm (TBFPA)
22.3 Experiment
22.3.1 Test Functions and Parameter Settings
22.3.2 Experimental Results and Analysis
22.4 Conclusion
References
23 A Video Transmission Control Strategy in Wireless Network
23.1 Introduction
23.2 Video Transmission Control Strategy
23.2.1 Data Coding Strategy Based on LT Codes
23.2.2 Code Adaptive Coding Redundancy
23.3 Experimental Results
23.4 Conclusion
References
24 Data Correction-Based Data Fusion Method for Internet of Things Terminals
24.1 Introduction
24.2 Sensor Data Processing
24.2.1 Data Spatial-Temporal Correlation
24.2.2 Data Weighted Fusion
24.3 Simulation Results and Analysis
24.4 Conclusion
References
25 Research on Multi-Sensor Data Fusion Algorithm for Monitoring of Power Distribution Station
25.1 Introduction
25.2 Data Fusion Algorithm
25.2.1 Fusion Algorithm Base on Closeness
25.2.2 Algorithm Improvement
25.3 Simulation Test and Analysis
25.4 Conclusion
References
26 Analysis of QR Reordering Algorithm Based on Feedback Technology Optimization
26.1 Introduction
26.2 QR Index Reordering Algorithm
26.3 Improved Reordering Algorithm Using QR Indicator
26.4 Conclusion
References
27 Example Verification of a Safety Evaluation Model for Road Passenger Transportation
27.1 Introduction
27.2 Evaluation Index
27.3 Evaluation Model
27.3.1 Evaluation Method
27.3.2 Traditional Evaluation Model
27.3.3 Improved Evaluation Model
27.4 Comparison Calculation of Two Examples
27.4.1 Example 1: A Traffic Accident
27.4.2 Example 2: A Safe Transportation Process
27.4.3 Comparative Analysis
27.5 Conclusion
References
28 Development of Broadband Sensors Based on Fiber Optic Gyroscope
28.1 Introduction
28.2 Instrument and Method
28.3 Configuration of Space Grade Single Axis IFOG
28.3.1 ASE Light Source
28.3.2 Fiber Optic Coil
28.3.3 MMS Performance
28.4 Conclusion
References
29 Short-Term Wind Speed Prediction Based on Phase Space Reconstruction and Elman Network-Grey Wolf Optimization
29.1 Introduction
29.1.1 Background
29.1.2 Phase Space Reconstruction
29.2 Elman Network
29.3 Grey Wolf Algorithm
29.3.1 Rationale of Grey Wolf Algorithm
29.3.2 Description of Grey Wolf Algorithm
29.3.3 Grey Wolf Optimization Algorithm
29.4 Wind Speed Combination Predication and Analysis
29.4.1 Prediction Process
29.4.2 Error Evaluation Function
29.4.3 Simulation Experiment and Result Analysis
29.5 Conclusion
References
30 Modern-Driven Deep Learning for Belief Propagation LDPC Decoding
30.1 Introduction
30.2 Related Work
30.2.1 BP Decoding
30.2.2 Model-Driven Deep Learning
30.3 Neural BP Structure
30.4 Simulation Result and Analysis
30.5 Conclusion
References
31 Blind Recognition of Channel Coding Based on Deep Learning
31.1 Introduction
31.2 Problem Formulation
31.3 Deep Learning Method
31.3.1 LSTM Recognizer
31.3.2 CNN Recognizer
31.3.3 Training Process
31.4 Experimental Results
31.5 Conclusion
References
32 Influence and Optimization of Levenberg–Marquardt Algorithm on Dielectric Loss Measurement in Noisy Environment
32.1 Introduction
32.2 High-Order Sinusoidal Fitting Method Is Used to Calculate the Dielectric Loss
32.3 Levenberg–Marquardt Algorithm
32.4 FIR Narrowband Bandpass Filter
32.5 Conclusion
References
33 Research on Artificial Intelligence Optimization Based on Genetic Algorithm
33.1 Artificial Intelligence
33.2 Genetic Algorithm
33.3 Research on Artificial Intelligence Optimization Based on Genetic Algorithm
33.3.1 Robocode
33.3.2 Optimization of Default Strategies
33.4 Conclusion
References
34 Multi-channel Scheduling for Intelligent Information Collection
34.1 Introduction
34.2 Communication Model of the Main Station for Intelligent Information Collection
34.3 Algorithm
34.4 Simulation Experiment
34.5 Conclusion
References
35 Anti-interference Performance of Transmit Diversity Techniques for Satellite System
35.1 Introduction
35.2 Discussion on Overhead Analysis Method
35.3 Overhead Analysis of Transmit Diversity for PUCCH Format 3 in Satellite System
35.4 Simulations
35.5 Conclusions
References
36 Wireless Sensor Energy Harvesting and Management
36.1 Introduction
36.2 Requirements for Energy Harvesting
36.3 Analysis of Energy Harvesting Technology
36.3.1 Solar (Light) Energy Collection
36.3.2 Wind Energy Harvesting
36.3.3 Harvesting of Thermoelectric Energy
36.3.4 Collection of Sound Energy
36.3.5 Magnetic Energy Harvesting
36.3.6 Comprehensive Harvesting of Energy from Various Sources
36.4 Energy Management
36.5 Conclusion
References
37 Stacked Sparse Autoencoder-Based WLAN Fingerprinting Localization
37.1 Introduction
37.2 System Model
37.2.1 System Overview
37.2.2 Fingerprinting Localization Using SSAE
37.3 Experimental Setup and Results
37.3.1 Experimental Setup
37.3.2 Experimental Results
37.4 Conclusion
References
38 A Gridless Joint DOA and Range Estimation Method for FMCW MIMO Radar
38.1 Introduction
38.2 Radar Model
38.3 DOA Estimation
38.3.1 Atomic Norm Denoising
38.3.2 TLS-ESPRIT DOA Estimation
38.4 Range Estimation
38.5 Simulations and Experiments
38.6 Conclusion
References
Author Index
Recommend Papers
![Advances in Wireless Communications and Applications: Smart Communications: Interactive Methods and Intelligent Algorithms, Proceedings of 3rd ICWCA 2019 [1st ed.]
9789811556968, 9789811556975](https://ebin.pub/img/200x200/advances-in-wireless-communications-and-applications-smart-communications-interactive-methods-and-intelligent-algorithms-proceedings-of-3rd-icwca-2019-1st-ed-9789811556968-9789811556975.jpg)
![Advances in Wireless Communications and Applications: Wireless Technology: Intelligent Network Technologies, Smart Services and Applications, ... Innovation, Systems and Technologies, 307) [1st ed. 2022]
9789811934858, 9789811934865, 9811934851](https://ebin.pub/img/200x200/advances-in-wireless-communications-and-applications-wireless-technology-intelligent-network-technologies-smart-services-and-applications-innovation-systems-and-technologies-307-1st-ed-2022-9789811934858-9789811934865-9811934851.jpg)
![Advanced Intelligent Technologies for Industry: Proceedings of 2nd International Conference on Advanced Intelligent Technologies (ICAIT 2021) (Smart Innovation, Systems and Technologies, 285) [1st ed. 2022]
9811697345, 9789811697340](https://ebin.pub/img/200x200/advanced-intelligent-technologies-for-industry-proceedings-of-2nd-international-conference-on-advanced-intelligent-technologies-icait-2021-smart-innovation-systems-and-technologies-285-1st-ed-2022-9811697345-9789811697340.jpg)
File loading please wait...
Citation preview
Smart Innovation, Systems and Technologies 257
Lakhmi C. Jain Roumen Kountchev Bin Hu Roumiana Kountcheva Editors
Smart Communications, Intelligent Algorithms and Interactive Methods Proceedings of 4th International Conference on Wireless Communications and Applications (ICWCA 2020)
Smart Innovation, Systems and Technologies Volume 257
Series Editors Robert J. Howlett, Bournemouth University and KES International, Shoreham-by-Sea, UK Lakhmi C. Jain, KES International, Shoreham-by-Sea, UK
The Smart Innovation, Systems and Technologies book series encompasses the topics of knowledge, intelligence, innovation and sustainability. The aim of the series is to make available a platform for the publication of books on all aspects of single and multi-disciplinary research on these themes in order to make the latest results available in a readily-accessible form. Volumes on interdisciplinary research combining two or more of these areas is particularly sought. The series covers systems and paradigms that employ knowledge and intelligence in a broad sense. Its scope is systems having embedded knowledge and intelligence, which may be applied to the solution of world problems in industry, the environment and the community. It also focusses on the knowledge-transfer methodologies and innovation strategies employed to make this happen effectively. The combination of intelligent systems tools and a broad range of applications introduces a need for a synergy of disciplines from science, technology, business and the humanities. The series will include conference proceedings, edited collections, monographs, handbooks, reference books, and other relevant types of book in areas of science and technology where smart systems and technologies can offer innovative solutions. High quality content is an essential feature for all book proposals accepted for the series. It is expected that editors of all accepted volumes will ensure that contributions are subjected to an appropriate level of reviewing process and adhere to KES quality principles. Indexed by SCOPUS, EI Compendex, INSPEC, WTI Frankfurt eG, zbMATH, Japanese Science and Technology Agency (JST), SCImago, DBLP. All books published in the series are submitted for consideration in Web of Science.
More information about this series at https://link.springer.com/bookseries/8767
Lakhmi C. Jain · Roumen Kountchev · Bin Hu · Roumiana Kountcheva Editors
Smart Communications, Intelligent Algorithms and Interactive Methods Proceedings of 4th International Conference on Wireless Communications and Applications (ICWCA 2020)
Editors Lakhmi C. Jain KES International Selby, UK Bin Hu Changsha Normal University Changsha, Hunan, China
Roumen Kountchev Faculty of Telecommunications Technical University of Sofia Sofia, Bulgaria Roumiana Kountcheva TK Engineering Sofia, Bulgaria
ISSN 2190-3018 ISSN 2190-3026 (electronic) Smart Innovation, Systems and Technologies ISBN 978-981-16-5163-2 ISBN 978-981-16-5164-9 (eBook) https://doi.org/10.1007/978-981-16-5164-9 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd. The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721, Singapore
Preface
This book is the first volume of the papers presented at the 4th International Conference on Wireless Communications and Applications (ICWCA2020) which was carried out on December 18–20, 2020, at Hainan University, Sanya, China. The papers cover the contemporary areas of smart communications, intelligent algorithms and interactive methods and related applications and present the latest achievements of the authors. The approach of the authors is based in general on the analysis and processing of multiple signals and events data obtained from various networks and wireless sensors and performed by using a great diversity of up-to-date tools: big data analysis, neural networks, contemporary architectures, algorithms, deep learning, and augmented reality. Special attention attracts the presented applications based on novel approaches in the areas, which give a reliable and convincing impact in the contemporary quality of life, such as the application of artificial intelligence in wireless sensor network coverage; design of cooperative driving systems; multi-user communication algorithm; intelligent rail-guided vehicle and computer numerical control; application of big data in management of distribution network operation; intelligent vehicle verification system; remote intelligent guidance system based on augmented reality; intelligent information collection, remote teaching, and many others.
v
vi
Preface
The aim of the book is to present the latest achievements of the authors to a wide range of readers, namely IT specialists, engineers, physicians, Ph.D. students, and other specialists. Selby, UK Sofia, Bulgaria Changsha, China Sofia, Bulgaria June 2021
Lakhmi C. Jain Roumen Kountchev Bin Hu Roumiana Kountcheva
Acknowledgments The book editors express their special thanks to chapter reviewers for their efforts and goodwill to help for the successful preparation of the book. Special thanks for Prof. Dr. Lakhmi C. Jain (Honorary Chair), Prof. Dr. Srikanta Patnaik (Program Chair), Prof. Chong Shen and Prof. Kun Zhang (General Chairs), Prof. Dr. Madjid Tavana (International Advisory Chair), and Silai Zhou (Organizing Chair) of ICWCA2020. The editors express their warmest thanks to the excellent Springer team which made this book possible.
Contents
1
Algorithm and Application of Artificial Intelligence in Wireless Sensor Network Coverage . . . . . . . . . . . . . . . . . . . . . . . . . . . Ruiyang Gao, Tingran Chen, Xin Zhang, and Ziliang Lv
2
Design of Cooperative Driving System Based on VANET . . . . . . . . . . Jing Tao Li, Xiao Jiang Li, Zhen Feng Li, and Run Qiu
3
A Study of Intelligent Power Information Collection System Based on Distributed Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bin Sun, Haiying Li, Ziping Long, Hua Zhang, Zhiliang Zhao, Huaizhen Jiang, and Lirong Pu
1 7
13
4
A Multiuser Communication Algorithm for UAV Bee Combat . . . . . Li Shi-ming, Yang Yuan, Wu Yi-jie, and Kang Chao
21
5
Review on Dec-POMDP Model for MARL Algorithms . . . . . . . . . . . . Shen Guicheng and Wang Yang
29
6
Research on AGV Correction Control Method Using QR Code Positioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yu Bin, Zhang Yi, Ye Chao, and Xing Bin
37
Research on Multi-data Center Collaboration Technology for Multi-station Fusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hu Song, Quan Fang, and Fei Xia
45
A Novel Method for Online Control in Equipment Parameters of Blow Molding Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Zhisong Zhu, Xu Zhao, and Yu Liu
53
Research on the Optimum Ratio of Intelligent Rail Guided Vehicle and Computer Numerical Control . . . . . . . . . . . . . . . . . . . . . . . Man Li, Xuewei Huang, Jiaqi Wang, Haiyun Yue, and Chaolang Hu
63
7
8
9
vii
viii
Contents
10 An Apple Grading Method Based on Improved VGG16 Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Lin Kaiyan, Zhang Haoping, Liu Chang, Si Huiping, Wu Junhui, and Chen Jie
75
11 Comprehensive Application of Big Data in Lean Management of Distribution Network Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jieming Zhang, Weiwen Li, Yuqian Chu, and Yanzhi Liang
87
12 Research on the Construction of Intelligent Vehicle Verification System for Road Transportation . . . . . . . . . . . . . . . . . . . . . Haiying Xia, Shuquan Xu, Yingji Liu, Xinlei Wei, and Hong Jia
97
13 Quantitative Risk Analysis of Dangerous Goods Road Transport Vehicles and Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 Xuan Dong, Zhong Xu, and Li Yuan 14 Virtual Reality Technology Facilitates Customized Clothing Design in C2M Business Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 Sun Jian 15 Research on Identification Model of Special Transformer Stealing Electricity Based on Vector Similarity Matching Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 Xin Xuan, Yazhuang Cao, Shuming Wang, and Tingyan Jiang 16 International Checkers Evaluation Algorithm Based on Evolutionary Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 Zhenghao Pan, ShuQin Li, and Shijing Gao 17 Research on Intelligent Health Diagnosis of Tunnel Structure System Based on Wavelet Neural Network . . . . . . . . . . . . . . . . . . . . . . . 137 Ruijun Li and Yaqing Shi 18 Data on Cloud Practice Based on Alibaba Cloud IOT Platform . . . . 143 Lei Wang 19 Remote Intelligent Guidance System Based on Augmented Reality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 Kaijia Luo, Zhenxiang Li, Zhiqi Guo, Yanan Zhang, and Zhang Lu 20 Research on Air Engine Air Circuit Fault Based on Convolutional Neural Network Feature Extraction . . . . . . . . . . . . 159 Yi Ren Zhou, Zhi Qiang Guo, and Chao Zhou 21 Improved Adaptive Particle Swarm Optimization Algorithm with a Two-Way Learning Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171 Yawen Wang, Qian Qian, Yong Feng, and Yunfa Fu
Contents
ix
22 Flower Pollination Algorithm Based on Beetle Antennae Search Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181 Jiayi Zhou, Qian Qian, Yunfa Fu, and Yong Feng 23 A Video Transmission Control Strategy in Wireless Network . . . . . . 191 Run Zhou, Min Xiang, and Yu Liu 24 Data Correction-Based Data Fusion Method for Internet of Things Terminals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199 Rujie Lei, Min Xiang, Peng Wang, Ruiheng Ma, and Kun Yu 25 Research on Multi-Sensor Data Fusion Algorithm for Monitoring of Power Distribution Station . . . . . . . . . . . . . . . . . . . . 207 Peng Wang, Min Xiang, and Rujie Lei 26 Analysis of QR Reordering Algorithm Based on Feedback Technology Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215 You Xinhua, Long wenjia, and Dai Yu 27 Example Verification of a Safety Evaluation Model for Road Passenger Transportation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223 Fujia Liu, Chaozhou Chen, and Guoliang Dong 28 Development of Broadband Sensors Based on Fiber Optic Gyroscope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233 Bo Yang, Yong Li, Yafei Zhao, Peiyong Zhang, Li Sun, Xin Hui, and Xiaona Zhou 29 Short-Term Wind Speed Prediction Based on Phase Space Reconstruction and Elman Network-Grey Wolf Optimization . . . . . 241 Shengqing Li, Haotian Feng, and Xi Deng 30 Modern-Driven Deep Learning for Belief Propagation LDPC Decoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251 Yiduo Tang, Lin Zhou, Shuying Zhang, and Yan Chen 31 Blind Recognition of Channel Coding Based on Deep Learning . . . . 259 Shuying Zhang, Lin Zhou, Zizhuo Zhao, and Yiduo Tang 32 Influence and Optimization of Levenberg–Marquardt Algorithm on Dielectric Loss Measurement in Noisy Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267 Chao Yu, Ronghui Hu, Xiaoxiao Xu, and Suaisuai Zhao 33 Research on Artificial Intelligence Optimization Based on Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279 Jingyi Liu 34 Multi-channel Scheduling for Intelligent Information Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285 Xingchuan Bao, Jianhua Yang, Jian Yang, Lin Peng, and Min Xu
x
Contents
35 Anti-interference Performance of Transmit Diversity Techniques for Satellite System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293 Ya Wen and Xin Wu 36 Wireless Sensor Energy Harvesting and Management . . . . . . . . . . . . 299 Huanan Zhang and Feng Wang 37 Stacked Sparse Autoencoder-Based WLAN Fingerprinting Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307 Junhang Bai, Yongliang Sun, and Quan Zhang 38 A Gridless Joint DOA and Range Estimation Method for FMCW MIMO Radar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315 Jingyu Cong, Tao Lu, Zhang Bin, Xiaoli Jing, and Xianpeng Wang Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
About the Editors
Prof. Dr. Lakhmi C. Jain, PhD, Dr H.C., ME, BE (Hons), Fellow (Engineers Australia), is with the Liverpool Hope University and the University of Arad. He was formerly with the University of Technology Sydney, the University of Canberra and Bournemouth University. Professor Jain founded the KES International for providing a professional community the opportunities for publications, knowledge exchange, cooperation and teaming. Involving around 5,000 researchers drawn from universities and companies world-wide, KES facilitates international cooperation and generate synergy in teaching and research. KES regularly provides networking opportunities for professional community through one of the largest conferences of its kind in the area of KES. http://www.kesinternational.org/organisat ion.php His interests focus on the artificial intelligence paradigms and their applications in complex systems, security, e-education, e-healthcare, unmanned air vehicles and intelligent agents.
xi
xii
About the Editors
Prof. Dr. Roumen Kountchev, D.Sc. is with the Faculty of Telecommunications, Dept. of Radio Communications and Video Technologies—Technical University of Sofia, Bulgaria. He has 434 papers published in magazines and conference proceedings, 18 books, and 22 patents. A member of Euro Mediterranean Academy of Arts and Sciences; President of Bulgarian Association for Pattern Recognition; Editorial board member of IJBST Journal Group; Editorial board member of: Intern. J. of Reasoning-based Intelligent Systems; Intern. J. Broad Research in Artificial Intelligence and Neuroscience; Editor of books for Springer SIST series. Bin Hu is an engineer and a counselor at Changsha Normal University, ACM Member, IEEE Member, GCDF, he is the guest editor of EI-indexed journals or book series: International Journal of Reasoningbased Intelligent Systems (IJRIS), International Journal of Information Systems and Supply Chain Management (IJISSCM), International Journal of Information Systems in the Service Sector (IJISSS) and Smart Innovation, Systems and Technologies (SIST), the technical editor of International Journal of Computer and Communication Technology (IJCCT). He published many articles in high standard journals like Knowledge-based System (KS), Journal of Intelligent and Fuzzy Control (JIFS), Enterprise Information System (EIS), Financial Innovation (FI) and etc. Dr. Roumiana Kountcheva, S.R. is Vice President of TK Engineering. She got her PhD at the Technical University of Sofia and became Senior Researcher at TIE, in 1993. She has 183 publications and 5 patents, and presented 16 plenary speeches at international conferences and workshops. She is a member of IRIEM, IDSAI, IJBST Journal Group, and Bulgarian Association for Pattern Recognition. She is a reviewer of WSEAS conferences and journals, and of books for Springer SIST series.
Chapter 1
Algorithm and Application of Artificial Intelligence in Wireless Sensor Network Coverage Ruiyang Gao, Tingran Chen, Xin Zhang, and Ziliang Lv
Abstract Artificial intelligence algorithm with its powerful processing ability to solve existing problems has attracted the attention of many experts and scholars. It has been gradually combined with the infinite sensor network, and a lot of research has been done on the energy control positioning system, router selection, and so on. Due to the slow convergence of particle swarm algorithm is easy to be the best to solve these problems, this paper presents a particle swarm algorithm based on artificial potential field (virtual force particle swarm optimization algorithm, VFPSO), in the artificial potential field algorithm, particle swarm optimization algorithm of iterative process introduced virtual force between particles, and the initial solution of optimization problem is more evenly spread out, the front-end optimization iteration of repulsive force is stronger. In the later stage, the anchor nodes are enhanced to improve the convergence efficiency of the algorithm, and the best solution with higher quality is found.
1.1 Wireless Sensor Network Application and Coverage With the rapid development of micro-electromechanical system, wireless communication, information network, and integrated circuit, the development of wireless sensor network technology formed on this basis also advances by leaps and bounds. Most of the sensor nodes are installed by air drop and other selection methods, which are used for monitoring in a large range of monitoring areas or hostile parties or dangerous areas [1, 2]. Wireless sensor network (WSN) can be simulated as consisting of three modules: data information acquisition network, data information dispersion network, and control center, among which the main element is sensor node. The sensor node is generally composed of data information collection, data information disposal, data information transmission, and power supply. Infinite sensor nodes can be independent of the network setup, used for data collection at the same time by the routing protocol (such as dance by single or multiple relay routing) will be R. Gao (B) · T. Chen · X. Zhang · Z. Lv Beihang University, Beijing, China e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_1
1
2
R. Gao et al.
collected data transmission of information to the information data acquisition nodes (sink), data acquisition node and data disposal by the wide area network (such as the Internet or satellite network) to achieve communication with the remote center, to collect the information for treatment [3].
1.2 Overview of Particle Swarm Optimization Based on Artificial Potential Field Particle swarm optimization algorithm is a biological population activity behavior originated from birds. In the process of bird seeking for food, the whole bird group maintains the unity, and individuals also retain the common mechanism of information, so that the bird group will not be excessively gathered and scattered. Since the proposal of algorithm was issued, researchers have focused on optimizing the function and strategy of algorithm, and at the same time, it is widely used in many industries such as update problem, automatic control and information mining, etc. [4]. For a long time, many scholars have carried out a wide range of research and analysis on it and gained a lot of progress. The basic idea of artificial potential field method is to simulate the influence of fictitious force field on the robot during its movement, and the movement direction of the robot is the combination of the influence of the attraction of the target object and the exclusion force of the blocking object. The mobile robot is simplified as a particle in two-dimensional space, and the movement direction X (X = (X, y)) at any orientation in the operating space is the total field strength direction of the repulsive field of the object and the attraction field of the target. Because of its practical simplicity and high energy effect, artificial potential field method has been widely used in the robot path planning industry. The basic thinking of artificial potential field is that the robot in the path environment can judge its running track by perceiving the attractive potential field of the target point and the crowding force field of the obstructions. In the environment, Q represents the area where the object is located, QR represents the target area, U(Q) represents the simulated potential field, so the virtual gravitational potential field is represented as U att (Q), and the virtual repulsive potential field is represented as U rep (Q). The virtual potential field and virtual force received by the robot in a certain region are shown as follows (the virtual force can be obtained by differentiating the virtual potential field) U (q) =Uatt (q) + Urep (q) − → F (q) = − ∇U (q) = −∇Uatt (q) − ∇Urep (q)
1 Algorithm and Application of Artificial Intelligence …
3
Its U(q) represents the gradient of U at Q, which is a vector, where the direction represented is the direction with the greatest rate of change of potential field at Q. For position q of x and y in two dimensions, ∇U (q) =
∂U ∂x ∂U ∂y
The minimum distance from the robot to the obstacle is P(q). ξ /η is the proportional coefficient. The fictional power obtained is: −→ Fatt (q) = − ξ(q − qg ) 1 1 η −→ Frep (q) = 2 ∇ p(q) − p (q) p(q) p0 Finally, the sum can be calculated by the crowding force and attraction force accumulated on the robot, and the movement direction of the robot can be determined according to the sum force. Virtual force algorithm is an artificial potential field method. All nodes are treated as charged particles and there are interaction forces between nodes. When the distance between nodes is less than a certain threshold, repulsive forces are generated between nodes. When the distance between nodes is greater than the threshold, the attraction between nodes is generated. The virtual force accepted by a node is the resultant force generated by all nodes, and the resultant force is used as the direction to guide the node to move. Within the area coverage, nodes can be dispersed from a dense area more quickly to achieve a uniform distribution effect, as shown in Fig. 1.1. If S 1 is a mobile node, S 2 , S 3 , and S 5 are all within the communication range of S 1 ; there is a virtual force between these nodes and S 1 , and S 4 is not within the communication range, so there is no force on S 1 . If the neighbor nodes of S 1 , S 2 , and S 3 on S 1 Fig. 1.1 Schematic diagram of virtual force action
4
R. Gao et al.
repulsion F 12 and F 13 , S 5 and S 1 are far apart but produce less attractive F 15 , S 4 does not produce force, then S 1 by force of F 1, F 12 , F 13 , F 15 , the sum of the power of the mobile node S 1 along the resultant force direction from the current position to move to a new location. Simulation S = {S 1 , S 2 , S 3, …, S N } is node set, so node i and j are S i and S j , the distance d ij between them is as follows: di j = d(si , s j ) =
(xi − x j )2 + (yi − y j )2
Therefore, the force of node S j on S i is formulated as follows: ⎧ if, ⎪ ⎪ 0, − → ⎨ (wa (di j − Dthe ), αi j ), if, Fi j = ⎪ 0, if, ⎪ ⎩ (wr (Dsi − di j ), αi j + π, if, d ij aij ωa /ωr Dthe Rc
di j Rc di j di j
> Rc ≥ di j > Dthe = Dthe < Dthe
S j on S Euclidean distance. S j on S direction angle of line segment. a virtual force gravitational force repulsive force parameter. Set distance threshold. the communication radius of the node.
The resultant force exerted on the node is as follows: → − → − Fi j Fi = j∈S
In the process of particle swarm optimization, the gravitational force and repulsive force are introduced into the robot in the artificial potential field method, and the virtual force generated between the particles is used to enlarge the particle optimization range and prevent the particles from falling into the local optimal value. After initializing the particle position, the particle distribution is more uniform in order to prevent excessive agglomeration of particles.
1.3 Use of Intelligent Algorithms in WSN Coverage Many experts and scholars at home and abroad have realized the advantages of intelligent algorithms, and successfully applied them in data disposal, automatic control, prediction category, identification pattern, medical category, military category, and so on, and achieved good results. It is also widely used in the infinite sensor network to deal with many problems such as routing, positioning, coverage selection, and energy
1 Algorithm and Application of Artificial Intelligence …
5
control, among which the application of intelligent algorithm in WSN coverage is mainly studied [5]. Particle swarm optimization is a traditional intelligent optimization algorithm. Since its introduction, it has attracted the attention of many researchers. At the same time, it has been applied in many researches and has grown into a corresponding separate discipline, which is applied in intelligence improvement, information discovery and signal disposal. It is widely used in many fields such as automation control, flow shop management, and network transmission. Different from other intelligent algorithms, particle swarm optimization focuses more on comprehensive retrieval. In the process of iteration, particle swarm optimization can reach a certain category and finally obtain the optimal solution. It has a strong robustness that makes it easier to operate. In the optimization process, this algorithm is easy to be troubled by the local optimization when solving the problem with high complexity and fall into the iterative process of infinite loop, which does not hinder the retrieval of the optimal solution. The first factor that falls into the infinite loop is that the global optimal value and the historical optimal value of each particle in the iteration process will affect the update of particle position. Although the initial motion will be improved under the action of algorithm convergence, the later algorithm is difficult to break and seek for the optimal solution, so it is difficult to get rid of the trouble of local optimal. After studying and practicing particle swarm optimization algorithm and other intelligent algorithms, the limitations of the algorithm are obviously improved, the original defects of the algorithm are overcome, and the optimal solution or more perfect solution can be obtained through the optimized algorithm. Now integrating particle swarm optimization with other intelligent algorithms has become an important optimization method to improve particle swarm optimization.
1.4 Conclusion This paper improves the standard particle swarm algorithm, the first artificial potential field algorithm and the convergence of particle swarm optimization, and take charge of the interaction of the introduction of particles in the optimization process of thinking, try to use the optimal particle in the iteration process for the target point and exclude force generated between particles avoid algorithm falls into local optimum, and finally to seek the optimal solution. Particle swarm optimization promotes particle average dispersion, which can effectively avoid excessive particle convergence and make the algorithm fall into the local optimal solution. However, due to the large amount of calculated work, it is more suitable to solve the problem of small coverage and obtain a better layout. Due to its simplicity, particle swarm optimization algorithm is more suitable for solving large area coverage problems. Compared with standard particle swarm optimization algorithm, particle swarm optimization algorithm can obtain the optimal network layout more efficiently.
6
R. Gao et al.
References 1. Jiang, J.R., Chen, Y.C., Lin, T.Y.: Particle swarm optimization for charger deployment in wireless rechargeable sensor networks. Int. J. Parallel Emergent Distrib. Syst., 1–16 (2018) 2. Mahboubi, H., Aghdam, A.G.: Distributed deployment algorithms for coverage improvement in a network of wireless mobile sensors: relocation by virtual force. IEEE Trans. Control Netw. Syst. 4(4), 736–748 (2017) 3. Ma, F., Zhang, L., Wang, J.B.: Improvement of particle swarm optimization algorithm and its application in optimization function. Comput. Digit. Eng. 7, 1252–1293 (2017) 4. Frans, V.D.B.: An analysis of particle swarm optimizers. University of Pretoria (2002) 5. Jain, N.K., Nangia, U., Jain, J.: A review of particle swarm optimization. J. Inst. Eng. 3, 1–5 (2018)
Chapter 2
Design of Cooperative Driving System Based on VANET Jing Tao Li, Xiao Jiang Li, Zhen Feng Li, and Run Qiu
Abstract The rapid development of automobile industry has brought great convenience to people’s travel, but the increase of vehicles also brings a series of environmental problems and traffic problems, so this paper designs a cooperative driving system based on vehicular ad hoc network (VANET). As a product of the integration of communication industry and automobile industry, VANET can provide drivers with the perception ability beyond visual range, thus giving drivers more reaction time and reducing the probability of traffic accidents. In addition, in some specific environment, the system can achieve a certain degree of automatic driving, reduce the driver’s fatigue, but also reduce fuel consumption. Finally, this paper uses Simulation of Urban MObility (SUMO) to simulate the vehicle driving at the crossroads. According to the simulation results, under the action of the system, the vehicles can pass through the intersection safely and effectively.
2.1 Introduction Every year, many people all over the world die in traffic accidents. Most of these accidents are caused by the driver’s wrong judgment. With the birth of the vehicle auxiliary driving system, many sensors are installed on the body, which helps the driver to make decisions to a certain extent and reduces the incidence of traffic accidents. But only using the sensor, the sensing distance of the car is relatively close, so it cannot play a good role of the sensor. Sometimes, it may even cause the driver’s misjudgment because of the false alarm of the sensor. As a new technology, VANET has integrated automobile industry and communication. It can interconnect the vehicles on the road with vehicles, between vehicles and people and between vehicles and fixed nodes, to form an easy to deploy and cheap intervehicle communication network [1]. The network can effectively manage the traffic flow, make traffic more fluid, and reduce traffic congestion and the occurrence of the accident. Because of the characteristics of self-organization, no center, and J. T. Li (B) · X. J. Li · Z. F. Li · R. Qiu Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming 650051, China © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_2
7
8
J. T. Li et al.
supporting multihop forwarding of data, the application of vehicular ad hoc network in cooperative driving system can provide drivers with the perception ability beyond visual range, which can provide more time for drivers to deal with some emergency situations [2]. At the same time, it can also provide a series of functions such as traffic accident early warning, traffic congestion query, and path planning. In addition, because the vehicle is connected to the Internet, it can also provide more entertainment for the passengers on the vehicle.
2.2 Overall System Model The system is composed of road test system, vehicle system, and human–computer interaction interface. The overall structure of the cooperative driving system is shown in Fig. 2.1. The system is composed of road test system, on-board system, and human– computer interface. When the system is in operation, on the one hand, the road test system collects traffic lights within a certain range and at the same time counts the number of current road vehicles and numbers these vehicles and calculates the average speed and flow density of vehicles in the current area. On the other hand, it combines the height of the system precision map, real-time monitoring of road conditions, and then the combination of the two information sent to the road test control unit, and the road test control unit makes decisions according to the information obtained and judges the crowded section and accident emergency section
Fig. 2.1 Overall structure of the cooperative driving system
2 Design of Cooperative Driving System Based on VANET
9
Fig. 2.2 Structure diagram of the collaborative driving system
in the area. Finally, the decision-making information is broadcast to the vehicles in the area through dedicated short range communication (DSRC) [3]. The on-board system records the vehicle’s own speed, acceleration, angle, and other motion state information through the speed sensor and inertial sensor. At the same time, it displays the geographic location of the current vehicle, such as roads and lanes, in combination with the vehicle’s own high precision map and global positioning system (GPS) location information. The DSRC can also continuously share the information of vehicle perception between vehicles to improve the perception range of vehicles. Then, the information of vehicle motion state and the perception information of surrounding vehicles are sent to the on-board control unit for data analysis and decision-making, so as to judge whether the current vehicle changes lanes, accelerates, and decelerates. Finally, the decision-making information is transmitted to the human–computer interface through controller area network bus. The driver analyzes the decision-making and perception information sent by the road test system and on-board system to determine whether the current vehicle is changing lanes, accelerating, decelerating, or following. The network structure diagram of the collaborative driving system is shown in Fig. 2.2.
2.3 Working Principle of the Collaborative Driving System 2.3.1 Vehicle Lane Change If a vehicle wants to leave the current road, it first needs to perceive the surrounding environment and traffic conditions of the vehicle, then reduce the speed, and enter the turning lane. The steps of lane changing are as follows: (1)
When vehicles need to change lanes, vehicles broadcast lane change information and their own speed and other dynamic parameter information to
10
(2)
(3)
J. T. Li et al.
surrounding vehicles through vehicle ad hoc network. Other vehicles will judge the distance between the two vehicles after receiving the information and remind the vehicle driver to drive carefully if the distance is close. If the driver of the rear vehicle fails to respond to the warning message sent by the vehicle within the specified time, the vehicle will automatically take braking measures. If there is a traffic accident on the road, the vehicles in the ad hoc network will broadcast traffic accident news to nearby vehicles. At the same time, the location of the traffic accident will be located according to the road test unit and the place where the vehicle GPS signal last appeared, helping the rescue team to be able to reach the place where the accident occurred in time for rescue. In addition, the vehicles near the traffic accident can transmit the warning information to the farther vehicles through the DSRC multihop communication technology [4]. After receiving the warning message, the remote vehicles will try their best to find other routes to avoid driving into the traffic accident section.
2.3.2 Vehicle Following When a vehicle is driving on a highway, the speed of the vehicle tends to remain the same for a period of time, so the vehicle can often choose to follow the car, so that the fatigue of driving can be effectively resolved. The vehicle uses the high precision map and GPS information to calculate the distance between the vehicle ahead and the current vehicle. At the same time, proportion integral differential (PID) control algorithm is used to synchronize the speed of the current vehicle with the speed of the vehicle in front, so that the distance between the front and rear vehicles is always kept at a safe distance [5]. When the distance between the front and rear vehicles is less than the safe distance, the warning function can be activated. When the driver fails to realize the danger or ignores the warning information, it can automatically perform the brake function to ensure the driving safety. The following steps are as follows: (1)
(2) (3)
The vehicle obtains external environment information through its own perception, and information exchange in the vehicle ad hoc network uses high precision maps and GPS to detect the distance between the current vehicle and the vehicle in front and uses visual sensors and image processing technology to determine the lane of the vehicle in front. Through the vehicle sensor to obtain the current vehicle speed, acceleration and other driving state information and the driver’s operation status. According to the status information obtained, the car-following system will be withdrawn when the current vehicle or the current vehicle needs to drive out of the current road. At the same time, the risk assessment of collision between the front and rear vehicles is carried out. The assessment result will be sent to the early warning system and brake control system, and the driver will send out early warning information through the alarm when there is a risk of collision.
2 Design of Cooperative Driving System Based on VANET
11
Fig. 2.3 Simulation flowchart
2.4 System Simulation In all traffic accidents, the intersection is the place where traffic accidents happen most easily, so this paper uses SUMO simulation software to generate a crossroad intersection and generate a number of vehicles on the road. The simulation flowchart is shown in Fig. 2.3. The goal of simulation is to transmit the turning intention of each vehicle to other vehicles at the intersection through vehicular ad hoc network, so as to realize the safe and effective passage of vehicles through the intersection. The simulation results show that the cooperative driving system based on VANET can effectively improve the driving safety, and the simulation results are shown in Fig. 2.4.
2.5 Conclusion In order to solve the problem of vehicle driving safety, this paper designs a cooperative driving system based on VANET. The system is based on wireless ad hoc network, which expands the driver’s perception distance, and has a certain warning function for the coming traffic accidents, so as to give the driver more reaction time and avoid the occurrence of traffic accidents. At the same time, in some specific scenarios, the system can achieve a certain degree of automatic driving, so it can effectively
12
J. T. Li et al.
Fig. 2.4 Simulation results of the collaborative driving system
help drivers alleviate driving fatigue and reduce the probability of traffic accidents. Finally, this paper uses SUMO to simulate the vehicle driving at the crossroads. According to the simulation results, under the action of the system, the vehicles can pass through the intersection safely and effectively.
References 1. Dou, Z.X., Wang, X.N.: Data delivery in VANET with network gaps. J. High Speed Netw., 111–120 (2020) 2. Kumar, N., Kaur, K., Jindal, A., Rodrigues, J.J.: Providing healthcare services on-the-fly using multi-player cooperation game theory in Internet of Vehicles (IoV) environment. Digit. Commun. Netw., 191–203 (2015) 3. Raza, H., Ioannou, P.: Vehicle-following control design for automated highway systems. IEEE Trans. Control Syst. Technol., 44–57 (1996) 4. Van Dijck, T., van der Heijden, G.A.:Vision sense: an advanced lateral collision warning system. In: IEEE Proceedings, Intelligent Vehicles Symposium, pp. 290–302 (2005) 5. Jia, D.Y., Zhang, R., Lu, K.J., Wang, J.P.: Improving the uplink performance of drive-thru internet via platoon-based cooperative retransmission. IEEE Trans. Veh. Technol., 4536–4545 (2014)
Chapter 3
A Study of Intelligent Power Information Collection System Based on Distributed Architecture Bin Sun, Haiying Li, Ziping Long, Hua Zhang, Zhiliang Zhao, Huaizhen Jiang, and Lirong Pu Abstract In allusion to a series of flaws in conventional power information collection system such as weak compatibility between terminals, communication blindness of collection points, low level of acquisition system integration, non-synchronous data checking and processing, and poor security protection performance, this paper puts forward and builds an intelligent power information collection system based on a three-tier distributed architecture which has main station layer, communication layer, and acquisition layer. Its purposes are to realize the following functions: automatic collecting power information, abnormal monitoring, power quality monitoring, power consumption analysis and management, information distribution, distributed power supervision, information interaction between smart power equipment, etc. With features of technical advancement and economic rationality, this system has been proved to show a way to upgrade project of rural power grid of Yunnan Agricultural Reclamation Power Co., Ltd, Xishuangbanna Dai Autonomous Prefecture, Yunnan Province, China, and further create a basic development condition for construction of regional smart grid.
B. Sun (B) Yunnan State Farms Group Co. Ltd., Kunming 650051, Yunnan, China e-mail: [email protected] H. Li School of Foreign Languages, Yunnan Minzu University, Kunming 650504, Yunnan, China Z. Long · H. Zhang · Z. Zhao Yunnan Agricultural Reclamation Power Co. Ltd., Jinghong 666100, Xishuangbanna, Yunnan, China H. Jiang XJ Electric Co. Ltd., Xuchang 461000, Henan, China L. Pu Zhuhai Pilot Technology Co. Ltd., Zhuhai 519000, Guangdong, China © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_3
13
14
B. Sun et al.
3.1 Introduction Already known are a series of problems in traditional power information collection system in Refs. [1–4]. First, the compatibility between terminals is weak. Being designed to access various types of collecting terminals, communication resolution and operation management between terminals with different interface standards are problems hard to deal with for traditional systems. Most data of the existing power information collection system are directly extracted from the marketing business system, and the inconsistency of real-time data affects also the accuracy of data collection. Human-caused data distortion in the absence of conditions for collecting and copying data and problems of the inability to store and extract data effectively in the conditions for collecting and copying data are to be solved urgently. Second, the communication of collection points is blind. With the continuous expansion of power grid construction area, the number of terminals connected to the power user electricity information acquisition system is becoming larger and larger, and the amount of collected data is larger too, so to uploading and storing all the collected data in real time are necessary. Anyway, the large number of substations, lines, and substation areas corresponding to a large number of newly added new acquisition points cannot match the original system standard. Third, the level of acquisition system integration is low. The present electric energy information acquisition system meets the needs of automatic meter reading, load control, and distribution transformer monitoring, but the data interaction with marketing business module, dispatching automation module, and distribution network automation module is not realized. So, system integration is a key problem to be solved urgently. Fourth, the data checking and processing are non-synchronous. Data recognition, judgment, and processing delay under meter code, power, voltage, current, power quality, and abnormal condition events largely reduce the availability of real-time data. Fifth, the security protection performance is poor. The system collects a large amount of data; simultaneously, the data backup, archiving, disaster recovery, and other relative strategies are particularly important. Effective strategic security management, real-time data, historical data testing, cost control function of low-voltage meter reading data, and terminal side data statistics and analysis are also included. What is clear from the above is that solving some existing problems in traditional system is imperative and creating a new system architecture is urgent. Based on discussions in Refs. [5–7], this paper aims to build an intelligent power information collection platform based on distributed architecture with features of much better customer satisfaction and power sense.
3 A Study of Intelligent Power Information Collection …
15
3.2 The System Architecture 3.2.1 The System Configuration The intelligent power consumption information acquisition platform includes master station layer, communication layer, and acquisition layer in Refs. [4, 8–14]. Its logical framework is shown in Fig. 3.1. The master station realizes data processing, analysis, storage, exchange, release, and comprehensive application platform functions; communication network, with the function of supporting multiple communication modes, realizes data transmission between terminal and master station. The acquisition layer realizes the functions of data acquisition, storage, processing, uploading, and user power consumption control.
Fig. 3.1 Logical framework of intelligent power consumption information acquisition platform
16
B. Sun et al.
3.2.2 The Master Station Layer The logic framework of the master station is constructed based on the core that contains power consumption information collection, data management and data sharing, and the focus, i.e., the application function of meeting users’ demand. Therefore, the unified database platform, data acquisition platform, data management platform, and comprehensive application platform are realized. The main station system functions include data acquisition management, business management, power management, application analysis, terminal management, and alarm management. Detailed functions are as follows. Data collection management includes implementation and current data, historical daily data, historical monthly data and event records, data rationality check, data analysis and calculation, and data storage management. The business management functions include meter reading function, meter change processing, current transformer (CT)/potential transformer (PT) replacement, bypass replacement, and manual test. Power management includes orderly power consumption management, power-fixed value control, electric quantity-fixed value control, rate-fixed value control, and remote control. Application analysis includes power statistics, load analysis, power quality data statistics, line loss and variation loss analysis, terminal online statistics, and report management. Terminal management includes basic data, terminal parameters, meter/user data, system timing, operation status management, maintenance, and fault recording. Alarm management includes alarm implementation, alarm processing, alarm event statistics, and alarm short message service (SMS).
3.2.3 The Communication Layer In this paper, access point name (APN) network is selected out of the security and timeliness of regional power grid. The master station server is connected to the general packet radio service (GPRS) network of the mobile company through an access point name (APN) special line. The routers of the both sides are interconnected by private fixed Internet Protocol (IP) address, and generic routing encapsulation (GRE) tunnel is used between gateway GPRS support node (GGSN) and interconnection router of communication operation company. The subscriber identity module (SIM) card used for general packet radio service (GPRS) private network can only open the service bound with the dedicated access point name (APN) and not available for other services. After obtaining access point name (APN), all general packet radio service (GPRS) modules are assigned private fixed Internet Protocol (IP), respectively. The end-to-end encryption is used between general packet radio service (GPRS) network table and application front-end computer to prevent information leakage in the process of transmission. Both sides isolate filter Internet Protocol (IP) and port on firewall. The technical project designed in this paper is suitable for
3 A Study of Intelligent Power Information Collection …
17
the agricultural reclamation power grid and other local power grids which are of high-security requirements, many data points, and strict timeliness requirements.
3.2.4 Intelligent Terminal Power consumption information acquisition terminals are responsible for the power information collection, data management, two-way data transmission, and forwarding or executing control commands. The terminals have three components: distribution transformer monitoring terminal, low-tension concentrate meter reading terminal, and power consumption management terminal.
3.2.4.1
The Capabilities of Distribution Transformer Monitoring Terminal
The terminal includes sub-functions like communication, data acquisition, time record reporting, and the like. The communication function includes high-efficiency data compression encryption algorithm, short message service (SMS) communication, general packet radio service (GPRS)/code division multiple access (CDMA)/public switched telephone network (PSTN)/Ethernet communication; data acquisition function includes multi-function meter data acquisition, alternating current (AC) analog acquisition, direct current (DC) analog acquisition, pulse acquisition, and switch state acquisition; event recording and reporting function includes transformer operation condition monitoring, energy meter operation status monitoring, abnormal power consumption monitoring events, and terminal self-monitoring events.
3.2.4.2
The Capabilities of Low-tension Concentrate Meter Reading Terminal
The low-voltage centralized reading terminal has functions like communication, expansion, distribution transformer monitoring, and the like. The communication function includes high-efficiency data compression encryption algorithm, supporting communication ways of short message service (SMS)/general packet radio service (GPRS)/code division multiple access (CDMA)/public switched telephone network (PSTN)/Ethernet. The expansion function includes carrier meter reading, automatic relay, automatic meter finding, and line loss calculation as well.
18
B. Sun et al.
3.3 Implementation of Intelligent Power Consumption Information Acquisition System 3.3.1 The Physical Architecture In this paper, the physical architecture based on Browser/Serve can reduce the cost of construction and maintenance and simplify the management process, too. The clients are designed web browsers and mobile apps, with a friendly application interface and simple operation, being suitable for large scale and complex application requirements. The component technology enables the system to have good scalability and meet the needs of building large-scale system. By adopting the cross platform design, the middle layer component provides a wide range of heterogeneous database access and replication capabilities. Therefore, automatically calculating time of use price and step price and connecting marketing system with other external systems seamlessly are realized. At the same time, with application of adopts the security design of hierarchical authority management mode and digital certificate authentication mechanism, a physical structure of intelligent electricity consumption information acquisition system is built based on Browser/Serve.
3.3.2 Internal and External Contact Normally, power grid enterprises give priority to the data acquisition system based on intrinsic safety since it meets the present needs of the current rural reclamation power grid architecture and the development requirements within a certain period in the future. The internal and external contact information of the system must ensure the safe transmission of information and risk-free storage of data.
3.4 Conclusions The system proposed in the paper has three layers: master station layer, communication layer, and acquisition layer. It solves the problems of poor compatibility of storage terminal, blind spot of communication coverage, low integration of acquisition system, asynchronous data verification and processing, poor system security protection performance and so on. It has been proved that the system is advanced in technology and economic in Xishuangbanna, Yunnan. It can also serve as a reference for the design and construction of intelligent power information collection system under the framework of smart grid. Acknowledgements This work was supported by the Golden Phoenix Talents Project in Yunnan State Farms Group Co., Ltd.
3 A Study of Intelligent Power Information Collection …
19
References 1. Sun, Q.Y., Yang, L.X., Zhang, H.G.: Smart energy-applications and prospects of artificial intelligence technology in power system. Control Decis. 33(5), 938–948 (2019) 2. Dileep, G.: A survey on smart grid technologies and applications. Renew. Energy 146, 2589– 2625 (2020) 3. Mohandes, B., Hammadi, R.A., Sanusi, W., Mezher, T., Khatib, S.E.: Advancing cyber–physical sustainability through integrated analysis of smart power systems: a case study on electric vehicles. Int. J. Crit. Infrastruct. Prot. 23, 33–48 (2018) 4. Ansari, J., Gholami, A., Kazemi, A.: Multi-agent systems for reactive power control in smart grids. Int. J. Electr. Power Energy Syst. 83, 411–425 (2016) 5. Sun, B., Li, H.Y., Guo, H., Li, T.: A study of smart system of power utilization safety management based on a cloud platform. In: Proceedings of the IACAE, pp. 1446–1450 (2019) 6. Sun, B., Sui, W.L., Li, H.Y.: Applied research of supervision and control system in 110 kV smart substation based on three layers of three networks. In: Proceedings of the ICIA, pp. 896–900 (2015) 7. Chen, J.C., Li, W.T., Wen, C.K., Teng, J.H., Ting, T.: Efficient identification method for power line outages in smart power grid. IEEE Trans. Power Syst. 29(4), 1788–1800 (2014) 8. Gong, J., Zhou, S., Niu, Z.S.: Optimal power allocation for energy harvesting and power grid coexisting wireless communication system. IEEE Trans. Commun. 61(7) (2013) 9. Didier, G., Lévêque, J., Rezzoug, A.: A novel approach to determine the optimal location of SFCL in electric power grid to improve power system stability. IEEE Trans. Power Syst. 28(2), 978–984 (2013) 10. Mohamed, O.A., Lutz, L.: Power line communications for low-voltage power grid tomography. IEEE Trans. Commun. 12, 5163–5175 (2013) 11. Menke, J.H., Bornhorst, N., Braun, M.: Distribution system monitoring for smart power grids with distributed generation using artificial neural networks. Int. J. Electr. Power Energy Syst. 113, 472–480 (2019) 12. Guo, Y., Huo, R., Liu, X., Xie, Z.: Lumped parameter cable model for low voltage power line communication. Autom. Electric Power Syst. 43(2), 156–161 (2019) 13. Alam, Z., Khursheed, A., Chaudhary, R.K.: Modeling simulation and performance evaluation of low voltage power line communication channel. Int. J. Adv. Technol. Eng. Explor. 5(46), 308–317 (2018) 14. Saleem, U., Jangsher, S., Qureshi, H.K., Hassan, S.A.: Joint subcarrier and power allocation in energy harvesting-aided D2D communication. IEEE Trans. Ind. Inf., 99 (2018)
Chapter 4
A Multiuser Communication Algorithm for UAV Bee Combat Li Shi-ming, Yang Yuan, Wu Yi-jie, and Kang Chao
Abstract The combat of unmanned aerial vehicle (UAV) bee colony is the primary trend of information warfare in the future. Under the high-dynamic environment of air-to-air combat, severe Doppler shifts generated by relative movement of UVA reduce the communication reliability between vehicles. Therefore, a new divisiontime multiuser chirp modulation (DT-MCM) algorithm based on chirp signal is proposed in this paper aiming to the characteristic of existing communication technology sensitive to Doppler shifts. Utilizing the insensitivity advantage of chirp signal to Doppler shift, two chirp signals with different chirp rate can constitute one chirp signal. And, our proposed method differentiates users using the chirp rate of chirp signals to implement the UAV bee colony communication. Depending on the electromagnetic environment changes, the efficient and reliable information transmission can be realized by dynamically adjusting parameters such as time and bandwidth. Applied to UAV bee colony communication and according to MATLAB simulation results, the proposed algorithm proved the effectiveness to reduce the adverse impact to UAV internetwork communication caused by Doppler shifts.
4.1 Introduction Unmanned aerial vehicle (UAV) is a kind of driverless aircraft using radio remote control device and its own program-controlled apparatus [1]. UAVs are playing an important role in recent year’s wars due to their small size, lightweight, flexible response, high-risk work performability, and effective reduction of combat personnel casualties [2]. As a subversive force to change the rules of future war, UAV bee colony has become the focus of research of domestic and foreign scholars [3, 4]. However, one of the key issues of UAV bee combat is the reliability between information transmission. When UAV bee colony flying, the nodes of the flight formation are close to and interact with each other as a whole. If an uncontrollable failure occurs L. Shi-ming · Y. Yuan (B) · W. Yi-jie · K. Chao Shanghai Aerospace Electronic Technology Institute, Shanghai 201100, China e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_4
21
22
L. Shi-ming et al.
at a bee colony node, it will inevitably affect other nodes, even collide. In air-to-air combat scenario, the relatively high-speed movement between vehicles will result in a large Doppler frequency shift, which will affect the colony communication [5]. Therefore, it is of great important to achieve the robust control of formation flight. In order to better improve the reliability of information transmission between UAV networks in high-dynamic scene, this paper selects chirp signal for communication, which is suitable for transmission in selective fading channels. Chirp signal has preferable robust to Doppler frequency and multipath ability [6] due to its good time– frequency characteristics. This letter proposes a novel multi-access communication algorithm based on chirp signal between UAV networks.
4.2 Chirp Signal Against Doppler Frequency Shift 4.2.1 Chirp Signal Impulse Compression Characteristic One of the significant features of chirp signal is good impulse compression [7]. When it performs matched filtering with its matched signal, the output signal can be written as: sin π Bt 1 − |t| √ T cos(2π f 0 t) S(t) = BT (4.1) π Bt where S(t) represents the output signal, B represents the signal bandwidth, and T is symbol duration. As shown in above formula, when chirp signal matched filtering, the output signal has similar performance to the sinc function which has a sharp autocorrelation peak. The time-domain waveform after matching modulation is shown in Fig. 4.1.
4.2.2 Analysis of Chirp Signal Anti-Doppler Performance This section analyzes the influence of the Doppler frequency shift on the impulse compression ability of the chirp signal. Assuming that, there is only Doppler shift in the channel, and the signal arriving at the receiving end after transmission through the wireless channel is expressed as: S(t) = cos(2π( f 0 + f d )t + π μt 2 )
(4.2)
4 A Multiuser Communication Algorithm for UAV Bee Combat
23
Fig. 4.1 Chirp matched filtering impulse compression waveform
where f d represents Doppler shift and μ is chirp signal frequency modulation index. After matched filtering received signal, the matched output signal is obtained as: r (t) = cos 2π( f 0 + f d )t + π μt 2 ∗ 2μ cos 2π f 0 t − π μt 2 √ sin(π (μt + f d )(T − |t|)) cos(2π ( f 0 + f d )t) = μ π (μt + f d )
(4.3)
It can be seen from formula (4.3) that when there is a Doppler shift in the channel and when the chirp signal matched filters with its matched signal, the output signal still has a similar performance to sinc function. However, the amplitude and peak position of the output waveform change, and the amplitude value decreases to 1 − | μfd | times. Since the chirp signal has a good time–frequency coupling characteristic, it can convert the frequency shift caused by the existence of Doppler into a time shift. The size of the time shift is μfd . Figure 4.2 is the overall diagram and the partial enlarged diagram of the output waveform of the chirp signal matched filter under different Doppler frequency shifts.
4.3 Division-Time Multiuser Chirp Modulation Algorithm From the above analysis, it can be observed that chirp signal has good anti-Doppler frequency shift performance. Therefore, based on the chirp signal, this paper presents a new type of communication algorithm suitable for UAV bee network, which can not only meet the reliable information transmission in high-dynamic air-to-air combat
24
L. Shi-ming et al.
Fig. 4.2 Chirp matched filtering impulse compression waveform under different Doppler shifts
scenarios but also save the system frequency band resources. This algorithm allocates user’s modulation frequency according to the time, so it is called divisiontime multiuser chirp modulation (DT-MCM). The time–frequency diagram of this algorithm is shown in Fig. 4.3. The algorithm uses chirp-BOK modulation. When the frequency modulation polarity is positive, namely μ > 0, data 1 is sent, and when the frequency polarity is negative, namely μ < 0, data 0 is sent [8]. Each user in the cellular network is assigned a pair of unique chirp signals composed of two frequency modulations, which are divided into two equal parts in a unified frequency band. The overall mathematical expression of the DT-MCM algorithm is as follows: ⎧ ⎨ s 1 (t) = 2E cos(2π f c t + π μ1 t 2 ), 0 ≤ t ≤ tk k k T sk (t) =
2E 2 2 2 ⎩ s (t) = cos 2π( f c + B 2)(t − tk ) + π μk (t − tk ) tk ≤ t ≤ T k T (4.4) Fig. 4.3 DT-MCM time–frequency diagram
4 A Multiuser Communication Algorithm for UAV Bee Combat
25
Fig. 4.4 Cross-correlation coefficient of DT-MCM algorithm
where k = 1, 2, 3, . . . , M represents the number of users in bee colony network, M is the total number of users, and E is the total energy of the signal in duration T. kT indicates the duration of each chirp rate of each user in network, and T tk = (k+1) and B are the duration of signal transmission and system bandwidth, respectively. μ1k represents the frequency modulation of the kth user in time 0 ≤ t ≤ T2 , and μ2k represents the complementary frequency modulation of the kth user in time 0 ≤ t ≤ T . The chirp rate of two periods is defined as follows: 2 (M + 1) B 2k T
(4.5)
(M + 1) B 2(M + 1 − k) T
(4.6)
μ1k = μ2k =
Regardless of the interference of non-cooperative parties, the key factor affecting the communication performance of bee colony network on the battlefield is the cross-correlation among network users [9]. Under ideal conditions, a good internetwork multiuser communication algorithm should ensure that the selection signals are orthogonal, that is, the cross-correlation coefficient among users is zero [10, 11]. Figure 4.4 displays the cross-correlation performance of the proposed algorithm drawn by MATLAB. Each signal is approximately orthogonal.
4.4 Simulation Results This section mainly verifies the accuracy of the previous theoretical derivation through MATLAB. Comparing with the existing communication algorithm DFMCM under Gaussian white noise channel and Doppler channel, proposed DTMCM’s pros and cons are analyzed. The simulation parameter settings are shown in Table 4.1.
26 Table 4.1 Simulation parameter setting table
L. Shi-ming et al. Value of D
Signal bandwidth B (MHz)
Symbol time T (µs)
Number of simulation data
50/500
20
2.5/25
400,000
4.4.1 Gaussian White Noise Channel It can be seen from Fig. 4.5a that in Gaussian white noise channel, when time bandwidth product D = 500, the performance of proposed DT-MCM algorithm is similar to that of the existing algorithm DF-MCM, which both ensure the reliability of user information transmission in the network and effectively suppress the interference among users. However, while the resources of communication systemare limited, that is, when the D value is only 50, as shown in Fig. 4.5b, when E b n 0 is equal Fig. 4.5 Bit error rate curve of the two algorithms under different D values
4 A Multiuser Communication Algorithm for UAV Bee Combat
27
to 18 dB, proposed DT-MCM theoretical bit error rate of the system can reach at 10−4 , yet the DF-MCM algorithm requires E b n 0 equal to 22 dB. So, the proposed DT-MCM algorithm has better performance, that is, 4 dB gains.
4.4.2 Doppler Channel In the simulation of air-to-air combat conditions, when there is a Doppler shift in the channel, the performance of two algorithms is simulated by MATLAB. By comparison, the bit error rate curve is shown in Fig. 4.6. Fig. 4.6 Bit error rate curves of two algorithms with different D values in Doppler channel
28
L. Shi-ming et al.
Figure 4.6a demonstrates that when time bandwidth product resources are sufficient, both algorithms have good anti-Doppler capabilities, but at the expense of the same power, the DF-MCM algorithm has better performance. However, when the communication resources are limited, that is, the value of D is greatly reduced, as shown in Fig. 4.6b, the performance of the DF-MCM algorithm is sharply descended, while the DT-MCM algorithm still has a good ability to resist Doppler shift. The performance gain is reduced by about 1 dB. Therefore, the DT-MCM algorithm has better anti-Doppler frequency shift capability.
4.5 Conclusions For the UAV bee colony network system in air-to-air high-dynamic scene, this paper uses chirp signal for communication and gives a new multiuser multiple access communication algorithm. This algorithm can resist the impact of Doppler frequency shift on the communication system to ensure the safe and reliable information transmission between UAV networks. The simulation examples given in the article illustrate the effectiveness of this method.
References 1. Papandreou-, A.: Applications in time-frequency signal processing. J. Res. Child. Educ. 27(2), 127–152 (2003) 2. Proakis, J.G.: Digital Communications, 3rd edn. McGraw Hill, Boston (2001) 3. Gupta, C., Mumtaz, T., Zaman, M.: Wideband chirp modulation for FH-CDMA wireless systems: coherent and non-coherent receiver structures. IEEE Int. Conf. Commun. 4, 2455–2459 (2003) 4. Cook, C.E., Bernfeld, M.R.: Radar signals: an introduction to theory and application. Artech House, MA (1993) 5. Ziomek, L.J., Chamberlain, S.: Underwater acoustics: a linear systems theory approach. Phys. Today 40(12), 91–92 (1987) 6. Ji, H., Kim, W., Shim, B.: Pilot-less sparse vector coding for short packet transmission. IEEE Wirel. Commun. Lett. 8(4), 1036–1039 (2019) 7. Ji, W., Wei, W., Liang, Q.: Compressive sensing-based signal compression and recovery in UWB wireless communication system. Wirel. Commun. Mob. Comput. 14(13), 1266–1275 (2015) 8. Pivato, P., Dalpez, S., Macii, D.: Performance evaluation of chirp spread spectrum ranging for indoor embedded navigation systems. In: IEEE International Symposium on Industrial Embedded Systems, pp. 1–4. IEEE (2012) 9. Khan, M.A., Rao, R.K., Wang, X.: Performance of quadratic and exponential multiuser chirp spread spectrum communication systems. In: International Symposium on Performance Evaluation of Computer and Telecommunication Systems, pp. 58–63. IEEE (2013) 10. Diet, A., Berland, C., Villegas, M., et al.: EER architecture specifications for OFDM transmitter using a class E amplifier. IEEE Microwave Wirel. Comp. Lett. 14(8), 389–391 (2018) 11. Alyan, E.A., Aljunid, S.A., Anuar, M.S., et al.: Analysis of theoretical and simulated performance of indoor optical wireless system based on CDMA technology. In: International Conference on Electronic Design (2017)
Chapter 5
Review on Dec-POMDP Model for MARL Algorithms Shen Guicheng and Wang Yang
Abstract With the various application and rapid development of deep reinforcement learning in multi-agent systems, the field of multi-agent reinforcement learning (MARL) continues to emerge with various practical application problems and corresponding solutions. Since single-agent decision-making tasks are difficult to meet the demands of complex application scenarios, most tasks need to be accomplished through cooperation or competition among multiple agents. This paper first introduces the characteristics of multi-agent systems, the basic concepts of reinforcement learning, and the methods for modeling multi-agent systems, then analyzes the strengths and weaknesses of two multi-agent reinforcement learning models for MARL algorithms in chronological order, which are communication-based learning model and collaborative-based learning model. Finally, based on the emerging opportunities and challenges in the field of MARL, possible research directions are proposed for the researchers’ reference.
5.1 Introduction In the early 1980s, the trial-and-error approach to the psychology of animal learning sparked a revival of reinforcement learning (RL) as one of the main lines in the historical development of RL, and this focus on trial-and-error learning intersected with two other main lines: Optimal control and temporal-difference learning gives rise to the modern field of RL [1]. There has been a lot of researches on the application of RL methods to solve complex single-agent decision-making tasks, but in complex application scenarios, it is not enough to rely only on the decision-making ability of a single-agent system to solve complex problems, and it is difficult to cope with the environmental variability of complex system tasks through predefined agent strategies [2]. Therefore, the application of RL to multi-agent systems can not only improve the decision-making ability to cope with environmental instability through independent learning by the agents themselves, but also improve the performance of S. Guicheng · W. Yang (B) School of Information, Beijing Wuzi University, Beijing 101149, China e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_5
29
30
S. Guicheng and W. Yang
multi-agent systems by learning from the interaction between agents to accomplish various complex cooperative and competitive tasks. Multi-agent reinforcement learning (MARL) can be thought as a system in which multiple agents interact in a public environment. Most of the MARL algorithms based on the Dec-POMDP model use the centralized training decentralized execution (CTDE) framework to solve the complex problems of MARL. During training, agents can share parameters for better policy learning, while during execution, agents perform collaborative tasks with a decentralized policy based on local observations, thus providing a good advantage in terms of partial observability and communication constraints in MARL systems, which is now a standard paradigm in multi-agent setups.
5.2 Multi-agent Reinforcement Learning 5.2.1 Background of MARL In many cases, agents are rarely independent systems, and they coexist and interact with other agents in different ways. Thus, such systems consisting of a set of agents that may interact with each other are called multi-agent systems (MAS). The decisionmaking function of an agent is divided into perception and action. The control of MAS is divided into centralized and decentralized, but most MASs are based on decentralized control, which makes the agents’ decisions more dependent on themselves. Decentralized control is more fault-tolerant and robust than centralized control and is more suitable for large MASs. In addition, communication between agents in MAS is a two-way process, i.e., all agents may be both sender and receiver of messages [3]. Reinforcement learning (RL) is rooted in the study of animal behavior under the influence of external stimuli, generating behavior through direct interaction with the environment and optimizing behavior through repeated attempts. RL algorithms can be divided this way: policy-based, value-based, and actor-critic combining the first two methods. The policy-based method, also known as policy gradients, directly outputs for the agent the probabilities of the various actions to be taken next. The value-based method calculates the value of all output actions for the agent’s next step and then selects the action with the highest value. Actor-critic has two networks, namely the actor network and the critic network. Most of the MARL algorithms are based on deep reinforcement learning (DRL), which is the replacement of the traditional tabular RL approach with deep neural networks [4]. There are not many MARL algorithms based on policy gradients alone, as they have significant limitations. In contrast, there are more types of actor-critic MARL algorithms, such as the deep deterministic policy gradient (DDPG). The MARL algorithms adopted in the literature summarized in this paper are mainly extensions of both DQN and actor-critic algorithms.
5 Review on Dec-POMDP Model for MARL Algorithms
31
5.2.2 Modeling for MARL Problem A complication with MARL is that each agent makes decisions that affect the environment differently, and the difficulty is that each agent has no way of knowing exactly what the other agents are doing, and this uncertainty poses a significant challenge for cooperation between agents. To address the above uncertainties, we can model the MARL problem as a Dec-POMDP. Both MDP and POMDP models are used to model single-agent problems, while for multi-agent problems, we typically use Dec-POMDP model. In Dec-POMDP, agents are only aware of their own actions and do not observe each other’s actions, and since agents can only access their individual actions and observations in this model, belief states need to be maintained and updated based on joint actions and observations, i.e., a set of joint policies needs to be found to maximize the expected cumulative rewards.
5.3 Advances in MARL Algorithms 5.3.1 MARL Algorithm of Communication-Based Learning Most works on communication-based MARL assume the use of predefined communication protocols, rather than attempting to learn communication protocols. The communication-based learning MARL approach works by assuming the explicit existence of information interactions between agents, this information is generated by each agent based on its own local observations, and during learning and training, agents are asked to determine which agents around them to collaborate with based on the information generated, and decisions are made through these explicit messaging and interactions during execution. Jakob et al. [5] proposed two algorithms: reinforced inter-agent learning (RIAL) and differentiable inter-agent learning (DIAL). The proposed method is the first to use deep neural networks for differential communication in discrete communication channels and to use RL to successfully learn communication protocols in a complex environment involving sequences and raw input images. In the same year, Sainbayar et al. [6] proposed the CommNet algorithm. Unlike RIAL and DIAL, CommNet passes messages as continuous variables rather than discrete variables, where each agent observes the state locally and sends continuous communication messages to other agents. Singh et al. [7] have proposed a personalized IC3Net algorithm for controlled continuous communication. IC3Net controls continuous communication through a gating mechanism and sets personalized rewards for each agent to improve performance and scalability. Jiang et al. [8] proposed an attention-based communication model (ATOC) to allow agents to learn when they need to communicate and how to integrate shared information for cooperative decision-making.
32
S. Guicheng and W. Yang
Because of the simplicity of the ATOC approach to establishing communication groups, Daewoo et al. [9] introduced the medium access control (MAC) protocol from the communications domain into MARL, which led to the SchedNet algorithm. In the same year, Das et al. [10] also proposed an algorithm called targeted multiagent communication (TarMAC) for learning to communicate in a multi-agent setup. TarMAC allows each individual agent to actively select other agents to send messages to, and this targeted communication behavior is achieved through signature-based Soft Attention. The following year, Ding et al. [11] proposed the individually inferred communication (I2C) algorithm based on TarMAC, which enables agents to learn a priori knowledge of inter-agent communication. I2C is a priori network that learns a priori knowledge by performing causal inference through feedforward neural networks, so that the necessity of communication between agents can be accurately grasped.
5.3.2 MARL Algorithm of Collaborative-based Learning Unlike communication-based learning MARL algorithms, collaborative-based learning MARL algorithms do not require agents to negotiate by learning explicit communication protocols, and they combine the idea of multi-agent learning with RL to form collaborative-based learning MARL algorithms. Such MARL algorithms can be divided into two categories: value-based approach and actor-critic approach. Value-based Approach. The basic idea of the value-decomposition networks (VDN) algorithm proposed by Peter et al. [12] is to train a joint-Q network centrally, and after training, each agent can decentralize its execution with a local Q-network obtained by cumulative summation and then decomposition of the joint Q-network. Afshin Oroojlooy Jadid et al. [13] made improvements to the VDN by proposing the monotonic value function decomposition (QMIX) algorithm. Unlike the VDN structure, QMIX imposes positive weights on both the joint and local Q-networks so that the joint action values are monotonic in each agent, and this monotonicity has the advantage of allowing the controllability of the joint action values to be maximized in policy-off learning and ensuring consistency between CTDE policies. VDN and QMIX are limited due to their structural constraints and thus can only handle a small fraction of the decomposable MARL tasks. Thus, Kyunghwan et al. [14] proposed the QTRAN algorithm for the factorial decomposition method of MARL. Compared to VDN or QMIX, QTRAN applies a more general approach to factorization, so it covers a broader class of MARL tasks. To solve the problem that QMIX cannot be explored effectively due to monotonicity constraints, Anuj et al. [15] proposed the multi-agent variational exploration network (MAVEN) algorithm. MAVEN introduces a potential space for hierarchical control with a mixture of value-based and policy-based. A hierarchical control policy is used to influence individual agent network parameters for effective exploration. Yang et al. [16] introduced a new idea to address the shortcomings of the previous algorithms by introducing a probabilistic framework derived from the determinant
5 Review on Dec-POMDP Model for MARL Algorithms
33
point process (DPP) of quantum physics, which is used to describe the quality and diversity of sampling subsets from the underlying set. Furthermore, the authors show that Q-DPP can generalize solvers such as VDN, QMIX, and QTRAN; i.e., these methods can be derived from Q-DPP as special cases. Actor-critic Approach. Foerster et al. [17] proposed a COMA algorithm based on the actor-critic structure for solving the multi-agent belief states assignment problem in the Dec-POMDP problem. COMA’s innovation is the use of counterfactual baselines that marginalize the behavior of individual agents while ensuring that the behavior of other agents is fixed. Ryu et al. [18] proposed an algorithm with generative cooperative policy network called generative cooperative policy network (GCPN). The GCPN uses a decentralized greedy strategy network to generate greedy behavior during training and then a network of generative cooperative strategies to generate action samples, thus allowing other agents to explore the strategy space more efficiently. Mao et al. [19] introduced attention in the MADDPG algorithm and proposed the ATT-MADDPG algorithm to solve the problem that multiple agents are unable to model dynamic strategies due to their constantly changing learning strategies. ATTMADDPG enhances the capabilities of the centralized critic network with attention.
5.4 Challenges and Research Directions for MARL Algorithms The MARL algorithm still faces several challenges: 1. The goals are not unique: Due to the complexity of multi-agent tasks, there may be conflicts between the different goals of different agents; 2. Non-stationarity: As each agent improves its strategy according to its own revenue, the environment faced by each agent becomes unstable; 3. Scalability: As the number of agents grows, the joint action space will be exponential growth, leading to a dimensional disaster problem for the MARL algorithm. Thus, there are still some research directions that need to be further explored and developed among future MARL research work in the field of DRL. Deep MARL Theory: With the application and development of DRL in MAS, many MARL algorithms employ deep neural networks to solve MARL problems with scalability. It is expected to extend these theories to the problem setting of deep MARL and strengthen the theoretical development of deep MARL. Safe MARL: The main approach to safe RL is based on the concept of risk introduced under optimal conditions and regulating the exploration process to avoid risky behavior. Thus, learning with provable safety guarantees is not an easy task, and the area of safe MARL is a worthwhile direction of research. Model-based MARL: Most RL algorithms are model-free; i.e., they are constantly learning by trial-and-error in the absence of a specific environment model. There is relatively little literature on model-based MARL algorithms. However, with the development of model-based RL in recent years, model-based RL algorithms can
34
S. Guicheng and W. Yang
outperform model-free RL algorithms to some extent in some cases, so the extension of model-based RL to MARL to improve its sampling efficiency is a worthwhile direction of research.
5.5 Conclusion This paper briefly summarizes the literature on MARL algorithms based on the DecPOMDP model in the last five years, including the relevant background, algorithm principles, model framework, and research progress of such MARL algorithms. We summarize the MARL algorithms based on the Dec-POMDP model in two separate categories according to the chronological order of publication in the literature, the relevance of the algorithm, and the type of work, namely, communication-based learning and collaborative-based learning MARL algorithms, and the latter of which is subdivided into value-based and actor-critic methods. In the concluding part of the article, the research directions of the MARL work are also discussed. With the continuous development of deep reinforcement learning, MARL technology is becoming more and more mature; thus, it can be predicted that MARL technology will provide more intelligent and efficient solutions for multi-agent systems facing large and complex uncertainty environments, thus further promoting the development of artificial intelligence technology.
References 1. Richard, S.S., Andrew, G.B.: Reinforcement Learning: An Introduction, 2nd edn. The MIT Press Cambridge, Massachusetts London, England (2018) 2. Busoniu, L., Babuska, R., DeSchutter, B.: A comprehensive survey of multiagent reinforcement learning. IEEE Trans. Syst. Man Cybern. Part C (Appl. Rev.) 38(2), 156–172 (2008) 3. Nikos, V.: A Concise Introduction to Multiagent Systems and Distributed Artificial Intelligence. Morgan & Claypool, Ronald Brachman (2007) 4. Oroojlooyjadid, A., Davood, H.: A review of cooperative multi-agent deep reinforcement learning (2019) 5. Jakob, N.F., et al.: Learning to communicate with deep multi-agent reinforcement learning. arXiv preprint arXiv:1605.06676 (2016) 6. Sainbayar, S., et al.: Learning multiagent communication with backpropagation. In: Conference on Neural Information Processing Systems, 2016, pp. 2244–2252. Barcelona, Spain (2016) 7. Singh, A., Tushar, J., Sainbayar, S.: Learning when to communicate at scale in multiagent cooperative and competitive tasks. In: International Conference on Learning Representations 2019. New York University (2019) 8. Jiang, J.J.H., et al.: Learning attentional communication for multi-agent cooperation. In: Neural Information Processing Systems, pp. 7254–7264. Montréal, Canada (2018) 9. Daewoo, K., et al.: Learning to schedule communication in multi-agent reinforcement learning. In: International Conference on Learning Representations, 2019. Daejeon, South Korea (2019) 10. Das, A., et al.: TarMAC: Targeted multi-agent communication. In: International Conference on Machine Learning, 2019. Long Beach, California (2019)
5 Review on Dec-POMDP Model for MARL Algorithms
35
11. Ding, Z.L., Huang, T.J., Lu, Z.Q.: Learning individually inferred communication for multiagent cooperation. arXiv preprint arXiv:2006.06455 (2020) 12. Peter, S., et al.: Value-decomposition networks for cooperative multi-agent learning. arXiv preprint arXiv:1706.05296v1 (2017) 13. Tabish, R., et al.: QMIX: monotonic value function factorisation for deep multi-agent reinforcement learning. In: International Conference on Machine Learning, 2018. Stockholm, Sweden (2018) 14. Kyunghwan, S., et al.: QTRAN: learning to factorize with transformation for cooperative multiagent reinforcement learning. In: International Conference on Machine Learning, 2019. Long Beach, California (2019) 15. Anuj, M., et al.: MAVEN: Multi-agent variational exploration. In: Conference on Neural Information Processing Systems, 2019. Vancouver, Canada (2019) 16. Yang, Y., et al.: Multi-agent determinantal Q-learning. In: Proceedings of the 37th International Conference on Machine Learning, 2020. Vienna, Austria (2020) 17. Foerster, J., et al.: Counterfactual multi-agent policy gradients. national conference on artificial intelligence, pp. 2974–2982 (2018) 18. Ryu, H., et al.: Multi-agent actor-critic with generative cooperative policy network. arXiv preprint arXiv:1810.09206 (2018) 19. Mao, H., et al.: Modelling the dynamic joint policy of teammates with attention multi-agent DDPG. arXiv preprint arXiv:1811.07029v1 (2018)
Chapter 6
Research on AGV Correction Control Method Using QR Code Positioning Yu Bin, Zhang Yi, Ye Chao, and Xing Bin
Abstract Aiming at the problem of displacement and angle deviation during the operation of automated guided vehicle (AGV), this paper uses quick response (QR) code as the guide landmark and proposes a visual-aided positioning method and AGV movement trajectory correction algorithm. First, the QR code is read and recognized by an industrial camera, the position of the QR code feature point is extracted, and the position and angle deviation parameters between the AGV and the QR code landmark are calculated, and then the trajectory correction algorithm is used to control the AGV to perform local curve movement. The AGV motion trajectory is corrected. Finally, the operation test is carried out in an indoor environment. The test results show that the combined use of visual positioning and trajectory correction algorithms can reduce the cumulative error caused by inertial navigation, avoid the AGV from being coded, and improve the operation efficiency of the AGV.
6.1 Introduction Automated guided vehicle (AGV) is a key equipment in the automatic logistics transportation and flexible manufacturing system. With the development of the logistics industry and the improvement of technology and technology, it is more and more widely used [1]. AGV can realize the handling task, and navigation and positioning play a key role. At present, the navigation methods of AGV mainly include electromagnetic navigation, optical navigation, laser navigation, inertial navigation, and visual navigation [2]. At present, most AGVs use laser navigation and magnetic navigation. These navigation technologies have disadvantages such as high cost and large straightness error of navigation tape [3]. Considering the factors of cost and positioning accuracy, Y. Bin (B) · Z. Yi · Y. Chao College of Advanced Manufacturing Engineering, Chongqing University of Posts and Telecommunications, Chongqing, China X. Bin Chongqing Innovation Center of Industrial Big-Data Co. Ltd, Chongqing, China © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_6
37
38
Y. Bin et al.
the quick response code (QR code) navigation technology came into being, and the current position information is obtained through the on-board camera scanning to guide, so as to realize the accurate positioning and navigation of the AGV.
6.2 AGV Correction Model Establishment The motion structure of AGV mainly includes four categories: wheel type, foot type, crawler type, and trunk type [4]. The AGV structure in this article uses a square wheel structure. There is a driving wheel at each end of the horizontal center axis of the car body, which is controlled by two DC motors. A universal wheel is installed at the four corners to maintain the balance of the car body and achieve steering function. Secondly, in order to solve the problem of wheel slippage of the AGV during the driving process, the AGV in this paper uses coaxial encoders on both driving wheels to measure the true distance of the AGV and provide accurate physical parameters for the positioning of the AGV [5]. The differential AGV completes linear motion, in situ rotation, and curved motion through the differential motion of the driving wheels on both sides of the body. The radius of the curve movement is determined by the speed difference V of the driving wheels. The AGV kinematics model is shown in Fig. 6.1. The AGV shown in Fig. 6.1 makes a curved movement with a radius of R. According to the kinematic analysis, we can get: V =
Vl + Vr 2
(6.1)
Among them: V is the speed of the center point of the vehicle body, and V l and V r are the speeds of the left and right driving wheels, respectively. Fig. 6.1 AGV motion model
6 Research on AGV Correction Control …
ω=
39
(Vl − Vr ) D
(6.2)
ω is the angular velocity of the curve movement, and D is the wheel distance between the two driving wheels; from expression (6.1) and expression (6.2), the radius R of the curve movement can be obtained: R=
V D(Vl + Vr ) = ω 2(Vl − Vr )
(6.3)
6.3 AGV Pose Correction System 6.3.1 QR Code Deviation Parameter Calculation In this paper, the accuracy of local QR code positioning is used to correct the pose of the AGV. When the AGV arrives at each node, the camera determines the position of the feature point of the QR code in the camera scan frame, and thus establishes a coordinate system to calculate the displacement and the angle deviation parameter. The definition of QR code image and deviation parameters is shown in the figure, where θ is the deflection angle of the AGV body and S is the displacement deviation value between the body and the QR code. The QR code deviation parameter definition diagram is shown in Fig. 6.2. Fig. 6.2 QR code deviation parameter definition diagram
40
Y. Bin et al.
The box in Fig. 6.2 is the scanning range of the camera, the side length is u, the black point is the center of the AGV vehicle body, and the coordinates of the four points A, B, C, and D are (ax , ay ), (bx , by ), (cx , cy ), and (d x , d y ). Then:
by − ay θ = arctan bx − ax
(6.4)
The slope of the QR code center line is cosθ, and the point p((ax + bx )/2, (ay + by )/2) is the point on the center line. From the point slope equation, the center line equation can be obtained as: ax + bx ay + by y = − cot θ x − + 2 2
(6.5)
Putting x = u/2 into expression (6.5), the ordinate of the intersection point F between the center line and the straight line x = u/2 can be obtained as: Fy =
ay + by u ax + bx − cot θ − 2 2 2
(6.6)
Then, the displacement deviation S of the AGV is: S = sin θ
u 2
− Fy
(6.7)
6.3.2 Research on Correction Algorithm of AGV Motion Track The motion trajectory is corrected for the displacement and angle deviation generated during the operation of the AGV, so that the AGV can be accurately positioned directly above the QR code when it travels to the next QR code landmark. From the previous deviation parameters, it can be seen that the AGV needs to be corrected in the following six states. When S > 0, there are three cases: θ > 0, θ = 0, and θ < 0; when S < 0, there are also three cases: θ > 0, θ = 0, and θ < 0. According to the motion characteristics of the dual-driving-wheel AGV, the AGV deviation can be attributed to the following three situations in Fig. 6.3. For the above three cases, this paper adopts the control of two drive wheel differential driving for one or two sections of arc motion to eliminate the displacement and angle deviation. S > 0, AGV need through as shown in Fig. 6.4 three trajectory to adjust deviation. When θ < 0, the AGV corrects position and angle errors by traveling in two symmetrical circular motions, which can be calculated from geometry knowledge.
6 Research on AGV Correction Control …
41
Fig. 6.3 AGV body deviation
Fig. 6.4 AGV correction movement trajectory
R=
sin |θ | × S 2
(6.8)
When θ = 0. √ S 2 + L 2 R= L 4 cos arctan S
(6.9)
where L is the distance between two QR codes. When θ > 0. R=
L sin |θ |
(6.10)
6.4 AGV Operation Experiment The AGV built for this experiment is shown in Fig. 6.5. The experiment is carried out indoors, and the QR code is evenly pasted on the ground at a distance of 50 cm. Applying the AGV correction algorithm to the AGV control program, the camera
42
Y. Bin et al.
Fig. 6.5 AGV of this experiment
and encoder will read the physical information and send it to the computer in real time to facilitate the monitoring of the AGV’s operating status. Figure 6.6 shows the running trajectory of the AGV in this article. It can be seen from Fig. 6.6 that the correction algorithm in this article can enable the AGV to independently correct the Fig. 6.6 AGV running track
6 Research on AGV Correction Control …
43
cumulative error generated during the operation, ensuring accurate positioning of the QR code, and the positioning error will not follow the operation. Increase with time and distance. The experimental results show that the error location and correction algorithm in this paper can be used in the correction control of AGV.
6.5 Conclusion Aiming at the error problem caused by the AGV during operation, the article proposes a visual error positioning method and an AGV motion trajectory correction algorithm. The QR code is applied to the AGV visual positioning, and the deviation parameters are calculated, and then the differential motion of the two driving wheels of the AGV is controlled, to correct the deviation. The experimental results show that the method proposed in the article can not only improve the positioning accuracy of the AGV, but also realize the self-correction through the trajectory correction algorithm, which greatly improves the operating efficiency. Acknowledgements Technology Innovation and Application Demonstration Project of Science and Technology Bureau of Beibei District, Chongqing (2020-5).
References 1. Yang, Q.M., Zhang, J., Ruan, Y.: Research on AGV trajectory planning and its correction control mechanism. Mech. Electr. Eng. 35(11), 1243–1247 (2018) 2. Wang, J., Xiao, X.Q.: Research on mobile robot compound navigation method based on QR code visual positioning. Chinese J. Sci. Instrum. 39(08), 230–238 (2018) 3. Liu, J.Y., Sun, W.T., Liu, M.: Research on path planning of automated guided vehicles in warehouse logistics. Mod. Mach. Tool Autom. Process. Technol. 6(12), 150–154 (2018) 4. Yu, J.: Application design of AGV system based on QR code technology in storage. J Tianjin Vocat. Tech. Normal Univ. 25(03), 44–47 (2015) 5. Zhang, K., Xu, L.H.: Research on magnetic navigation AGV differential steering control algorithm based on fuzzy control. Autom. Instrum. 31(10), 1–5 (2016)
Chapter 7
Research on Multi-data Center Collaboration Technology for Multi-station Fusion Hu Song, Quan Fang, and Fei Xia
Abstract This paper designs an intelligent collaboration mechanism between the distributed edge data center and the core data center platform. The core data center is responsible for managing and monitoring the distributed data center and the resources of the data center, coordinating the collaborative allocation of resources between the platforms to complete the needs of users. The distributed data center independently manages and monitors the computing resources of the data center and is responsible for processing specific storage and computing tasks. The platforms are connected to each other through the backbone network. This paper innovatively proposes a cloud edge resource graph intelligent scheduling method based on graph database technology that adapts to the collaborative operation of multiple data centers, adapts to multi-center resource collaborative operation scenarios, and is different from traditional centralized data center data resource scheduling methods. The bottom layer is based on graph database technology. Storing resource types and relationships between them, constructing a graph state library describing the correlation between multi-data center cloud-side collaborative resources, and using graph scheduling algorithms to convert heterogeneous multi-data center resource scheduling into directed acyclic graph scheduling problems. Using the minimum cut algorithm of the graph database and the algorithm based on the critical path ensures the optimal scheduling results in a complex multi-data center cloud-side collaborative resource scheduling environment.
7.1 Introduction With the advent of the internet of everything era, the amount of data generated by edge devices has increased rapidly, which has brought higher data transmission bandwidth requirements [1, 2]. At the same time, new applications have also put forward higher requirements for the real-time data processing. The traditional cloud computing model has been unable to respond effectively [3, 4]; edge computing has H. Song (B) · Q. Fang · F. Xia State Grid Jiangsu Electric Power Co., LTD. Infomation and Telecommunication Branch, Nanjing 211106, Jiangsu, China © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_7
45
46
H. Song et al.
emerged, and the coordinated development of cloud computing and edge computing will become an important trend in future development [5, 6]. The “multi-site integration” data center refers to the multi-level data center including the edge data center, which is used to store, process and calculate the data generated in the internal business of the power grid, urban governance, internet, industrial manufacturing and other fields. The gathering base of data information is also the main load center. Through the large-scale development of “multi-site integration” data center stations, the overall advantages of edge computing and cloud data centers of “business segmentation, cloud edge collaboration, and overall connection” can be created, which can significantly lead the development of the cloud data upstream and downstream industry chain and promote ubiquitous new business forms and new models of power internet of things appear. Multi-site integration inevitably has many sites. The grid existing data centers are mostly centralized centers, and the technology accumulation does not cover the coordination problems between multiple distributed centers. It is necessary to study the interaction mechanism and storage technology of resources between core stations and distributed stations to form a resource schedulable whole and exert the synergistic effect of several stations.
7.2 Resource Management Architecture of Distributed Data Center and Core Data Center Platform Based on graph database technology, this paper designs a cloud-side resource intelligent scheduling architecture that adapts to the collaborative operation of multiple data centers. It uses library technology to store heterogeneous resources and their relationships and builds a graph state library describing cloud-side collaborative resources in multiple data centers. The graph scheduling algorithm for multiple data centers achieves optimal scheduling results in a complex heterogeneous multi-data center cloud-side collaborative resource scheduling environment. The distributed data center and core data center platform resource management architecture is mainly composed of distributed data centers and core data centers deployed in multiple stations, and the platforms are connected to each other through a backbone network, as shown in Fig. 7.1. Difference from single data center, the core data center is responsible for managing and monitoring the distributed data center Fig. 7.1 Multi-data center resource management architecture diagram
7 Research on Multi-data Center Collaboration …
47
and its own data center resources, and coordinating the collaborative allocation of resources between the platforms to fulfill the needs of users. Each distributed data center independently manages and monitors the computing resources of the data center, and is responsible for processing specific storage and computing tasks. Different from single data center infrastructure service, this paper proposes a multi data center infrastructure service structure. This architecture mainly provides infrastructure services and meets the requirements for collaborative resource management under a distributed multi-data center architecture. Based on the design concept of distributed data centers, distributed data centers located in different geographical locations can satisfy real-time service requests from local users. The core data center platform and each distributed data center interact with each other through internal and external networks. The isolation mechanism of each distributed data center subnet ensures the secure network isolation of the resources of each data center. Different from the single data center resource management mode, this paper proposes the core data center platform master control center. The core data center platform master control center is the resource management center of the entire system. It manages and maintains the resource information and data information of all distributed data centers. After the control platforms of each distributed data center are deployed, they can immediately join the overall computing resource pool through registration and access authentication at the master control center. The master control center carries out the coordinated allocation of resources in multiple data centers according to user needs. Each data center is used as a management unit to obtain the resource usage status of each data center; each data center manages each physical server or computing cluster in the area as a unit. It can monitor the operation and load of each physical host in the area. Users register and authenticate the platform in the core data center through the terminal and can submit resource request tasks to the core data center according to their own needs. The core data center will formulate corresponding strategies and user levels according to user needs, select the appropriate service area and deliver tasks to the corresponding data center and return to the service address of the user data center. At this time, the user can directly access the service provided by the corresponding data center according to the service address, and the terminal will not need to interact with the core data center again during the process of using the cloud service, which increases the real-time nature of the service. The core data center needs to be connected to each data center. On the one hand, it can obtain the resource usage of each data center in time and synchronize the resource data information in each region in real time in the core data center; on the other hand, it can synchronize each data center. The center issues cloud resource tasks to meet various resource needs raised by users.
48
H. Song et al.
7.3 Multi-level Allocation and Recycling Technology for Resources in Multiple Data Centers Based on the resource management architecture, design multi-data center resource allocation and resource optimization recovery plan to realize the optimized management of distributed data center and core data center platform resources. The specific plan to be adopted is as follows in Fig. 7.2. Under the multi-data center resource management architecture, the core data center formulates resource allocation strategies according to user needs and collaborates to organize the computing and storage resources of each edge cloud. Resource management mainly includes the allocation and recycling of resources. The process of resource allocation in multiple data centers is as follows: Step 1: The user submits a request for resource application through the platform. The application information includes the area where the resource is located, the type of the resource operating system, and the resource configuration such as the number of CPU cores, memory size, hard disk size, etc.; Step 2: If the user selects local resources, the corresponding data center can create resource allocation tasks, otherwise the core data center needs to create resource allocation tasks in the master control center; Step 3: The execution of the task first selects the service area, and then the system forwards the task to the corresponding distributed data center; Step 4: The resources in the distributed data center adopt the method of polling scheduling to query the running status and resource usage of each physical resource in the data center. If there is no idle resource that meets the request, it will directly return the virtual machine creation failure errors and log records of the system. At the same time, the platform sends the user a notification that the resources in the area are unavailable, and the user re-selects the service area, that is, go to step 1; if an idle resource that meets the request is found, the corresponding physical resource is allocated resource of; Step 5: If the resource is created successfully, the platform generates the location information of the resource according to the actual situation, including the data center area ID number and resource ID number. Store resource location information and configuration information on the data storage server of the distributed
Fig. 7.2 Virtual and real fusion process diagram
7 Research on Multi-data Center Collaboration …
49
platform and perform data synchronization in the core data center control center in real time, and at the same time, notify the user of the successful resource allocation message and return the resource login information; Step 6: The resource allocation process ends. The process of multi-data center resource recovery is as follows: Step 1: The core data center and the distributed data center obtain the collection of resources that need to be recycled by monitoring the status of resources; Step 2: Query the location information of the resource to be recycled on the data storage server; Step 3: The administrator creates a resource recovery task; Step 4: To execute the task, first select the distributed data center indicated by the ID of the data center area where the resource is located and then forward the task to the corresponding distributed data center; Step 5: The distributed data center selects the corresponding physical resource or virtual resource in the area according to the server ID; Step 6: The distributed data center invokes the related methods of the scheduling server to close and recycle resources, and recycle useless resources to the total computing resource pool, and at the same time, update the resource data information in the distributed data center and the master control center and return message of successful resource recovery; Step 7: The resource recovery process ends.
7.4 Data Synchronization Technology for Collaborative Business Between Distributed Data Center and Core Data Center Platform The design of this paper is based on the multi-data center collaborative difference synchronization method, the design of a custom distributed graph database log format, the application of the data synchronization method based on the update log, and the design of a synchronization transaction plan to achieve transaction-based update and rollback data. In view of the inherent correlation of some update operations, once an error occurs during the synchronization process, simply rolling back or canceling an operation may affect the data in the unpredictable range, resulting in data inconsistency and synchronization failure. And this technology based on research can regard affairs as the basic unit, which improves the security and success rate of synchronization. The data synchronization method based on the update log includes event trigger and time trigger. In the data synchronization process of the two triggering methods, there may be errors such as data conflicts, transaction blocking, or concurrency. In order to avoid or solve this defect, in the synchronization process, the publishing user
50
H. Song et al.
and the subscribing user, respectively, adopt the TCP and HTTP transmission protocols, while designing the conflict detection and avoidance mechanism to ensure the integrity and correctness of the transmitted data during the synchronization process. Synchronous message transmission in distributed systems mainly adopts asynchronous mode and adopts group agent mechanism. The group agent mechanism uses a group of nodes to forward messages and increases the message transmission channels by adding message transmission agents in the network, and improves the reliability of message transmission. This project designs a multi-data center distributed publish/subscribe data transmission model for multi-site converged business rules. This model is oriented to multiple data centers based on a publish/subscribe communication model. Data transmission-related messages are all sent using a group proxy mechanism, making the publishing and subscription of application services more efficient and standardized, and data transmission more secure and reliable. This project is based on the group agent mechanism, combined with business rules, to design an appropriate distributed data center message publishing and subscription mechanism.
7.5 Conclusion This paper innovatively proposes a cloud-side resource graph intelligent scheduling method based on graph database technology that adapts to the collaborative operation of multiple data centers, adapts to multi-center resource collaborative operation scenarios, and is different from traditional centralized data center data resource scheduling methods. The bottom layer is based on graph database technology. Storing resource types and the relationship between them, constructing a graph state library describing the correlation between multi-data center cloud-side collaborative resources, using graph scheduling algorithm to convert heterogeneous multi-data center resource scheduling into directed acyclic graph scheduling problem, using the minimum cut algorithm of the graph database and the algorithm based on the critical path to ensure the optimal scheduling results in a complex multi-data center cloud-side collaborative resource scheduling environment. Acknowledgements This work is supported by Science and Technology Project of State Grid Corporation of China (Research and Application on Multi-Datacenters Cooperation and Intelligent Operation & Maintenance, No. 5700-202018194A-0-0-00).
References 1. Baker, J., Bond C.: Providing scalable, highly available storage for interactive services. In: Biennial Conference on Innovative Data Systems Research, pp. 223–234 (2011)
7 Research on Multi-data Center Collaboration …
51
2. Armbrust, M., Fox, A.: Griffith, R.: Above the Clouds: A Berkeley View of Cloud Computing. EECS Department, University of California, Berkeley (2009) 3. Manyika, J., Chui, M., Brown, B.: The next frontier for innovation, competition, and productivity, pp. 154–160 (2011) 4. Boettiger, C.: An introduction to Docker for reproducible research. ACM SIGOPS Oper. Syst. Rev. 49(1), 71–79 (2015) 5. Jo, M., Maksymyuk, T., Strykhalyuk, B.: Device-to-device-based heterogeneous radio access network architecture for mobile cloud computing. IEEE Wirel. Commun. 22(3), 50–58 (2015) 6. Armbrust, M., Fox, A., Griffith, R., et al.: A view of cloud computing. Commun. ACM, 53(4), 50–58 (2010)
Chapter 8
A Novel Method for Online Control in Equipment Parameters of Blow Molding Process Zhisong Zhu, Xu Zhao, and Yu Liu
Abstract A developed method for online control of equipment parameters in blow molding was proposed in this paper. It aimed to investigate the product quality fluctuation caused by the drift of parameters in the production process. The processing parameters and product quality data were got through orthogonal experiments. The product quality index predictor was established by random forest. The product quality indicators were monitored in real time. When the product quality indicators were out of the allowable range, the collaborative control strategy would adjust process parameters to return to the set value according to the quality indicator in time. In this paper, high density poly ethylene (HDPE) was used as the raw material, and the processing of intermediate bulk container (IBC) tank was taken as an example to verify the effectiveness of this method. The experimental results revealed that the developed control method can predict the change of product quality in real time and keep the product quality index to a reasonable range.
8.1 Introduction Hollow blow molding is a molding method that blows a hot-melt parison closed in a mold into a hollow product by means of gas pressure, which is one of the common plastic molding processing methods [1]. The traditional hollow blow molding quality inspection method is manual sampling and measuring the various indexes of the products at intervals. The feedback on the quality of the products has hysteresis, which leads to large change in weight and uneven wall thickness distribution [2, 3] of the hollow blow molding products. The influences of blow molding process parameters on the quality of blow-molded products are complex and nonlinear. The relationship between the blow molding process parameters and the quality of blowmolded products is difficult to be characterized by quantitative formulas. In order to solve the above problems and the fluctuation of some process parameters (process parameter drift) caused by environmental disturbances when producing a tank of Z. Zhu (B) · X. Zhao · Y. Liu School of Mechanical Engineering, Nantong University, Nantong 226019, Jiangsu, China e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_8
53
54
Z. Zhu et al.
1000 L intermediate bulk container (IBC) [4, 5], this paper proposed a method for online regulation of process parameters based on quality prediction model for which to achieve closed-loop control of product quality.
8.2 Product Quality Prediction Mode In this paper, the orthogonal experiment [6] was used to design 6 horizontal 10 factor orthogonal test; 26 sets of experimental data were studied. The process parameters of blow molding machine [7–9] related to quality index are selected, and the corresponding relationship between process parameters and quality indicators was got. Then a random forest quality model was established to evaluate 26 orthogonal experimental factors. It aimed to eliminate the process parameters which were not related to the quality of the blow-molded products during the blow molding process of plastic products and get 4 levels of 8 factor orthogonal data. The quantitative value of the correlation degree would be the weighting factor of the control parameters in cooperative control. The selected process parameters were training data to establish a random forest quality prediction model.
8.2.1 Orthogonal Design The main performance parameters of the blow molding machine included the upper die temperature, the medium die temperature, the lower die temperature [10], melt temperature [11], blowing pressure [12], pressure holding time [13], feeding length, screw speed, and partial temperature of two extruders (18 groups) [14]. Each of the identified test factors determined the respective number of levels based on actual conditions. The principle was to cover all ranges of values for all test factors. For example, the reasonable range of the pressure holding time is 120– 135 s. To cover all the values of the test factors and consider the principle of uniform distribution, there were 4 levels to test: 120, 125, 130, and 135 s. According to the actual production conditions, except for the feeding length to be set to 3 levels, the number of other factors is set to 6 levels. The process parameter levels were set as shown in Table 8.1.
8.2.2 Random Forest Prediction Model Random forest algorithm is an integrated algorithm [15]. In the model training, the random forest would evaluate the feature relevance to select the most critical process parameters for the quality index. The quantitative value of the correlation degree [16] would be the weighting factor of the control parameter in cooperative control.
8 A Novel Method for Online Control in Equipment …
55
Table 8.1 The process parameters level value Influencing factor
1
2
3
4
5
6
Upper die temperature (°C)
165
167
169
171
173
175
Medium die temperature (°C)
165
167
169
171
173
175
Lower die temperature (°C)
165
167
169
171
173
175
Melt temperature (°C)
170
174
178
182
186
190
Blowing pressure (Mpa)
0.54
0.56
0.58
0.60
0.62
0.64
Pressure holding time (s)
120
122
124
126
128
130
Feeding length (cm)
197
198
199
Screw speed (Hz)
40
41
42
43
44
45
No. 1 extruder temperatures 1–9 (°C)
150
154
158
162
166
170
No. 2 extruder temperatures 1–9 (°C)
150
154
158
162
166
170
According to the correlation analysis, 8 sets of parameters, such as melt temperature, blowing pressure, and feeding length accounted for 90% of the total importance. The feature variables were sorted in descending order of importance, and the attributes that accounts for 10% importance were deleted to get a new feature set. Based on it, a new random forest model was established, and the correlation degree of each feature was calculated in the feature set. The process parameters were listed according to the correlation degree: the upper die temperature, the medium die temperature, the lower die temperature, melt temperature, blowing pressure, pressure holding time, feeding length, and feeding speed. The test levels for each process parameter were listed in combination with expert experience and actual production as shown in Table 8.2. Table 8.2 The main process parameters level value Influencing factor
1
2
3
4
Upper die temperature (°C)
164
168
172
176
Medium die temperature (°C)
164
168
172
176
Lower die temperature (°C)
164
168
172
176
Melt temperature (°C)
180
185
190
195
Blowing pressure (Mpa)
0.54
0.58
0.60
0.62
Pressure holding time (s)
124
128
132
136
Feeding length (cm)
212
214
216
218
Screw speed (Hz)
40
42
44
46
56
Z. Zhu et al.
8.2.3 Process Parameter Correlation Analysis A quality predictor based on random forest regression was established by training experimental data got from orthogonal experiments. Among them, the inputs were the upper die temperature, the middle die temperature and the lower die temperature, melt temperature, blowing pressure, pressure holding time, feeding length, and screw speed. The outputs were the product weight and the wall thickness uniformity of the production (Wall thickness variance). The parameters of the random forest regression function based on scikit-learn [17] were set as follows: The maximum number of decision trees: n_estimators = 10; Evaluation criteria for characteristics: criterion = ‘gini’; The maximum depth of the decision tree: max_depth = None; The minimum number of samples required for internal node subdivision: min_samples_split = 2; The smallest sample weight of the leaf node is: min_weight_fraction_leaf = 0.0; Maximum number of leaf nodes: max_leaf_nodes = None; Whether to put back: bootstrap = True; The upper die temperatures, the middle die temperature and the lower die temper◦ ◦ ◦ ature were set: TM1 = 172 , TM2 = 168 , TM3 = 168 ; the melt temperature TR = 187◦ ; the blowing pressure P = 0.58 Mpa; the pressure holding time t = 130 s; screw speed V = 45 Hz; the feeding length L = 215 cm. The test results were shown in Fig. 8.1. Comparing the error between the test sample values and the quality estimates of the Fig. 8.1a–d, it can be found that the model generalization ability was remarkable. In Fig. 8.1a, the weight of the product decreased with the increase of the melt temperature when other process parameters were unchanged. In Fig. 8.1b, the wall thickness uniformity of the product decreased with the increase of the melt temperature when other process parameters were unchanged. In Fig. 8.1c, the weight of the product was not related to the blowing pressure under the same conditions of other process parameters. In Fig. 8.1d, the variance of the wall thickness uniformity of the product increased with the increase of the blowing pressure under the premise that other process parameters were unchanged. It showed that the quality prediction model got the relationship between the process parameters and product quality indicators. The random forest algorithm was used to fit the product quality and wall thickness uniformity, respectively. The correlation degree between the product quality and the process parameters was shown in Fig. 8.2. It can be seen that the feeding length, the melt temperature, lower die temperature, and the screw speed played major roles in the weight change of the product. Blowing pressure, melt temperature, lower die temperature, and pressure holding time played a major role in change of wall thickness uniformity correlation degree.
8 A Novel Method for Online Control in Equipment …
a
c
57
b
d
Fig. 8.1 a and b were the relationships between the melt temperature and the weight of the product, and the wall thickness uniformity of the production (wall thickness variance); c and d were the relationships between the lowing pressure and the weight of the product, and the wall thickness uniformity of the production (wall thickness variance) Fig. 8.2 The correlation degree between the product quality and the process parameters
58
Z. Zhu et al.
8.3 Process Parameter Collaborative Control During the blow molding process, fluctuations in certain process parameters (such as melt temperature and mold temperature) might be influenced by environmental factors (such as workshop temperature and humidity), which can produce deviation between the actual value and the set value of the process parameter. The resulted deviations would affect product quality; therefore, it was necessary to establish a closed-loop control system to ensure product quality.
8.3.1 Process Parameters Collaborative Control Specific Process The control strategy included: 1.
2.
The current process parameters were put into the quality prediction model, the model output the current quality indicators, and determined whether adjustment was needed; The relevant process parameters were adjusted according to the degree and direction of the correlation parameters of the process parameters. Ultimately, the quality index of the product was within the acceptable range. The specific operation process was followed:
Step 1: The weight index predicting the quality fluctuation value Y1i and the wall thickness uniformity index predicting the quality fluctuation value Y2i were calculated, the current parameters were put into the random forest quality model to predict the quality index, and judged whether it exceeds the index range by |Yki | > ek , ek was the quality index error range. The value of Yki is Yki = Qsk − Q ki , Qsk was the benchmark quality index, and it was the middle value of the quality index interval. Q ki was the current quality prediction value. Step 2: When both Y1i and Y2i did not exceed the error range of the quality index, it indicated that both quality categories met the quality index and continued to process according to the original parameters. When one of the indicators Y1i and Y2i exceeded the error range of the quality index, then the calculation was based on the parameters which exceeded the error range of the quality index and correlation degree corresponding to the indicators. When both Y1i and Y2i exceeded the error range of the quality index, the proportion of limit violations β1i and β2i were calculated and compared. If β1i > β2i , the current process parameters were adjusted according to the weight index predicting the quality fluctuation value Y1i . On the contrary, the current process parameters were adjusted corresponding to the wall thickness uniformity index predicting quality fluctuation value Y2i and the correlation degree.
8 A Novel Method for Online Control in Equipment …
59
Step 3: Adjust the current process parameters. The adjustment process is as followed: Calculate the adjustment amount of each iteration Ui j : Uki j = Ak j × Bi j × βki (1 + αki )
(8.1)
Ak j was the process parameter correlation quantified value, Bi j was the current process parameter, βki was the overrun ratio, its value was: βki = Yki /Qsk
(8.2)
Yki was the current predicted quality fluctuation value, Qsk was Benchmark quality indicator. αki was the adjustment speed, the size was: αki j = Yki − Yk(i−1) /Yki
(8.3)
Calculate the adjusted process parameters B(i+1) j : B(i+1) j = Bi j + Uki j
(8.4)
Determine if B(i+1) j was outside the process parameter range B j min , B j max . If it was not exceeded, the process parameters were in line with the quality index; if it is exceeded the range, the current process parameter category Bi j was unchanged. After the current process parameter category was removed, the remaining process parameter correlation quantified value Ak j was recalculated, Uki j and B(i+1) j were recalculated; Return to step 1 and iterate until the quality indicator was met. In the above process, k was the quality indicator category, j was the process parameter category, and i was the iteration count. The specific adjustment process was shown in Fig. 8.3:
8.3.2 Experimental Verification In order to verify the effectiveness of the process parameter collaborative control method, the melt temperature reduced by 8°to simulate a change in the situation of environmental. The data of product weight and wall thickness uniformity in each production cycle were recorded separately, and the change of product quality and wall thickness uniformity under coordinated control and without intervention were compared. The result was shown in Fig. 8.4.
60
Z. Zhu et al.
Fig. 8.3 A schematic diagram of the adjustment process
By comparison, the collaborative control method can find the combination of process parameters by the coordinated adjustment of process parameters, so the product quality can return to the qualified range and avoid the situation of a large number of unqualified products. The quality parameters after cooperative control had been in a reasonable range and the fluctuations were small.
8.4 Conclusion In order to solve the problem of process parameter drift in the actual production process of 1000 L intermediate bulk container (IBC) tank, this paper proposed an online regulation method for process parameters of blow molding process. The method monitored the quality of the product in time by establishing a random forest quality prediction model. When the product quality fluctuated too much, the process parameters of the product would be regulated by the collaborative control method indirectly. The results showed that the control method can monitor the quality changes of the products and restore the quality to a reasonable range in time. The method reduced process parameter drift and provided certain references for the online quality control of blow-molded products.
8 A Novel Method for Online Control in Equipment …
61
Fig. 8.4 The data of product weight and wall thickness uniformity in each production cycle were recorded separately, and the change of product quality and wall thickness uniformity under coordinated control and without intervention were compared
References 1. Au, K.M., Yu, K.M.: Conformal cooling channel design and CAE simulation for rapid blow mould. Int. J. Adv. Manuf. Technol. 6, 311–324 (2013) 2. Zeng J.S., Zeng, X.F.: Research on Wall Thickness Control of Parison by Hollow Blow Molding. China Plastics Ind. 09 (2010) 3. Musa, R.K, Tan, V., Kalyon, D.: Measurement and calculation of parison dimensions and bottle thickness distribution during blow molding. Polym. Eng. Sci. 21, 331–338 (1981) 4. Glor, M.: Requirements for flexible intermediate bulk containers (FIBC) current status of standardization. Sic Handhabung Brennbarer Sta. 2182, 163–174 (2012) 5. Nayan, P., Sunny, Zr., Himanshu, P.: Performance evaluation of HDPE/MWCNT and HDPE/kenaf composites. J Thermoplast. Compos. (2019.08)
62
Z. Zhu et al.
6. Deng, L., Feng, B.., Zhang, Y.: An optimization method for multi-objective and multi-factor designing of aceramic slurry. Combining orthogonal experimental design with artificial neural networks. Ceram Int. 44, 15918–15923 (2018) 7. Li, W.M., Xu, Y., Feng, D.Y., Zho, M.D.: Design and research of control system of the largescale accumulator blow molding machine based on PLC. Adv. Intell. Syst. Res. 23 (2012) 8. Wang, A., Li, X.J.: Application of BP neural networks PID control in blow molding machine system. Comput. Simula. 09 (2010) 9. Wang, L., Gao, Y., Xi, Z.: Application of computer technology in hollow blow molding machine. Plast. Sci. Technol. 10 (2015) 10. Ruan, Y.F., Gao, H., Li, D.Q.: Improving the consistency of injection molding products by intelligent temperature compensation control. Adv. Polym. Tech. (2019) 11. Biglione, J., Bereaux, Y., Charmeau, J.Y,. Balcaen, J., Chhay, S.: Numerical simulation and optimization of the injection blow molding of polypropylene bottles—a single stage process. Int. Mater. Form. 9, 471–487 (2016) 12. Zhao, C.H., Gao, F.R.: Melt temperature profile prediction for thermoplastic injection molding. Polyneng Sci. 39, 1787–1801 (2004) 13. Qiu, B., Li, W.: The effect of packing pressure and packing time on thickness distribute of injection moldingpart. Pla. Manuf. 11 (2006) 14. Du, Q.L., Chen, X.H.: Design and investigation of temperature control system for hollow blow molding extruding machine. CHINA Plas Ind. 07 (2006) 15. Breiman, L.: Random forests. Mach. Learn. 45, 5–32 (2001) 16. Huang, F.L., Lin, J.P., Lou, Y.J., Xu, J.H.: Research on robust design of craft parameters in plastic injection based on relation degree. CHINA Mech. Eng. 21 (2009) 17. Wang, Y.S., Xia, S.T.: A survey of random forests algorithms. Inf. Commun. Technol. 01 (2018)
Chapter 9
Research on the Optimum Ratio of Intelligent Rail Guided Vehicle and Computer Numerical Control Man Li, Xuewei Huang, Jiaqi Wang, Haiyun Yue, and Chaolang Hu
Abstract In the intelligent machining system, the ratio between the number of track-type automatic guide vehicle and the number of computer numerical control (abbreviation CNC will be used below) machine tool, as well as its scheduling plan have a decisive influence on the machining efficiency of the system. In this paper, we explore the optimum ratio of rail guided vehicle (abbreviation RGV will be used below) and CNC corresponding to the processing time of different individual parts under single process. Through dividing the processing time of a single part, the “2n step by step” model is designed for the RGV. And the traversal algorithm is used to find out the optimal path planning of RGV for all the CNC values in a given processing time. Then, by comparing the total waiting time of RGV and CNC in different cases, the optimum ratio of the certain processing time is determined, which provides a strong basis for the manufacturer on how to design start-up plan.
9.1 Introduction With the development of information technology, big data analysis and artificial intelligence technology, more and more problems of optimal scheduling and optimal control have emerged in many fields, such as industrial manufacturing, machining and logistics transportation. The problems have a wide range, such as material processing scheduling, job process scheduling, warehouse material storage scheduling, classification of logistics products and so on. RGV is a kind of unmanned intelligent car, which can run freely on fixed track. It can do simple repetitive work, reduce the waste of human resources and effectively improve the production efficiency of modern intelligent processing factory [1]. Thus, it is widely used in workshop and automatic three-dimensional warehouse. Its working range includes goods storaging, goods picking, AB conversion and other tasks. Particularly, RGV can cooperate with CNC to participate in loading, unloading and cleaning of a variety of parts that require different processing times. At this time, RGV and CNC form a processing system. M. Li · X. Huang · J. Wang · H. Yue · C. Hu (B) Math College of Sichuan University, Chengdu, China e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_9
63
64
M. Li et al.
Obviously, the working efficiency of the system is closely related to the number of RGV on track and the number of CNC, and the route planning of the system is also an important factor. Up to now, some articles have already studied the related problems of RGV. Zhang Guiqin and others [2] used collision avoidance algorithm to avoid the problem of two RGV colliding or trailing in the same direction; aiming at the influence on control and operation performance caused by time-varying uncertainty of sampling period in network control system. Shen Yan and others [3] proposed dynamic scheduling strategy based on feedback control principle and prediction mechanism. On this basis, Cha Zhenyuan and others [4] compared the usage of RGV and stacker under six scheduling strategies and determined the scheduling strategy with the highest utilization. Nevertheless, as mentioned above, the existing research results mainly focus on route design, and the ratio of RGV and CNC is mostly processed by a fixed value. But the optimal ratio between intelligent rail guide vehicle and CNC machine tool is seldom studied, and there is a lack of more general and universal conclusions. In fact, when there is a serious mismatch of the ratio between RGV and CNC, no matter how the scheduling scheme is designed, the production efficiency cannot be greatly improved. Because the RGV processing time and waiting time will become the bottleneck of the system. So we think when the ratio is fixed, the study on strategy scheduling has practical significance, and under the specified scheduling algorithm, according to the relative proportion relationship between the RGV processing time of a single product and the CNC processing time of a single product, studying the optimum proportion of RGV and CNC is equally important. Our research conclusions can provide a reference for manufacturers to determine how many CNC to start at the same time according to the CNC processing time of a single product, which will help to avoid the second change caused by unreasonable distribution of RGV and CNC quantity, as well as to save time and cost. Therefore, based on the CNC processing time of a single part, this paper establishes a RGV path planning model for each type of data, and optimizes the scheduling with the shortest sum of CNC waiting time in one shift (8 h) as the goal. Moreover, after the optimization scheme is obtained, the CNC optimal proportion determination model is established to find the optimal proportion and the corresponding scheduling scheme by minimizing the sum of the waiting time of RGV and CNC under the optimization scheme corresponding to each CNC value. The summary results provide the theoretical basis for the proportion setting in production practice, and at the same time, general conclusions are drawn, and the next research direction is explained.
9.2 Problem Analysis The working environment studied in this paper consists of a limited number of CNC, 1 intelligent rail guide vehicle, 1 straight track, 1 feeding conveyor belt, 1 blanking conveyor belt and other related auxiliary facilities. We stipulate that the CNCs are
9 Research on the Optimum Ratio of Intelligent Rail …
65
always placed by this rule: odd number on one side, and the corresponding even number on the other side. For example, when there are 8 CNC, the whole system is shown in the Fig. 9.1. Under the above working environment, each CNC is equipped with the same cutting tools for the same processing process. Each raw material must be processed on any CNC and cleaned by the RGV before it can be finished. This paper always considers the number of RGV to be 1 when discussing the optimal ratio of RGV and CNC corresponding to different processing time T, that is, only the number of CNC is changed. Since between the opposite two CNC, RGV does not need to spend extra moving time on loading material; it is considered that the actual production situation is more favorable when CNC appear in pairs. So in the follow-up discussion, the selection of CNC quantity is all in pairs. For avoiding unnecessary calculations and making the results more general, we first divide the processing time T into intervals. Since too long processing time is meaningless to the production mode discussed in this paper, combined with the actual production data, this paper discusses the processing time within 0–1200 s. First, the scheduling strategy of RGV with a fixed ratio of 1:n is discussed. In the process of optimizing the ratio, our ultimate goal is always to improve the processing efficiency of the system, that is, we hope that both RGV and CNC are in the working state as long as possible. Therefore, in a given processing time interval, we discuss the optimal path planning for RGV and then compare the sum of waiting time to select the optimum ratio corresponding to the processing time. When it comes to the problem of the optimal path planning, considering that when all the CNC are in the processing state, if RGV waited and not going anywhere after it completes the current instruction, unnecessary waiting time will be generated before any CNC sends signal to the RGV again, which will directly lead to the reduction of production efficiency. Therefore, when scheduling the RGV, it is stipulated that if the RGV does not receive any signal of feeding demand after completing the current instruction, the RGV will immediately discriminate according to the scheduling model and execute the moving instruction. As far as possible to increase the efficiency of the operation, this paper hopes to increase the number of finished materials in a shift, which means to make the CNC stay in work state as long as possible. In other words, the total waiting time of the CNC should be as small as possible. Since only local optimal solutions can be obtained for such problems, we consider designing a “2n step-by-step” scheduling model for RGV, where n is half of the number of CNC opened in the current planning. RGV will traverse 2n times to calculate the future movement (or stay-in-place) before discriminating the execution of the moving instruction, and it is required that each movement or stay-in-place must be separated by more than one instruction to move or stay here. After traversing the future movement for 2n times, RGV will take the first step of the path with the minimum total waiting time (or stay-in-place) as its next instruction. Based on the above calculation, when n changes from 1 to 10, according to the comparison of the total waiting time of RGV and CNC, we determine the optimal ratio of RGV and CNC for each case.
66
M. Li et al.
9.3 Model Building 9.3.1 Determination of RGV Movement Time Suppose that RGV moves at a constant speed, the time required to move 1, 2 and 3 units of RGV is obtained according to the data consulted. Further, the time required to move 0–10 units of RGV motion is obtained by fitting with the least square method. The fitting function is f = 16.6x + 1.6.
9.3.2 Estimation of CNC Upper Limit When RGV movement time is neglected ideally, the production of a finished material requires at least RGV loading, unloading and cleaning and one CNC processing. Since the loading and unloading time and cleaning time generally do not fluctuate greatly, and there is no substantial impact on the discussion of this problem, combined with the existing data, the cleaning time is taken as 27 s, numbered odd CNC the loading and unloading time is 28 s, numbered even CNC the loading and unloading time is 33 s. Therefore, the sum of loading, unloading and cleaning time for one product is 55 or 60 s. It is supposed that the total duration of a shift is T T = 8 h = 28800 s. So in the case of short processing time, RGV could process up to 525 finished materials within a shift. Accordingly, the CNC upper limit should be
1 525 ∗ 2 18800 1200
= 10(pairs)
(9.1)
Thus, for each range of processing time, we only consider the number of CNC between 1 and 10 pairs.
9.3.3 A Preliminary Division of Processing Time T for a Single Material Assuming that for a single process, the time of CNC machining single material is constant and consistent. Since too long processing time is meaningless to the production mode discussed in this paper, combined with the actual production data, this paper discusses the processing time within 0–1200 s. According to the data table of operation parameters of intelligent processing system, the processing time of CNC of a single material varies greatly under different specific conditions. To consider the problem comprehensively as far as possible under the condition of reducing the
9 Research on the Optimum Ratio of Intelligent Rail …
67
complexity, the partition interval is taken for 10 s. This kind of division is more representative and has strong operability in production practice.
9.3.4 The Optimal Path Planning Model Working principle of CNC. Assuming that in one shift (8 h), CNC working state is stable, that is, failure is not considered. There are four states of CNC: idle state, loading and unloading state, processing state and demand state in the whole course of working. The idle state represents the state of no material on the CNC; the loading and unloading state represents the state of the RGV is feeding the CNC; the processing state represents the state of the material being processed; and the demand state represents the state of waiting for the RGV blanking after the completion of the processing. Let the state of the CNC numbered i be Ai , and assign values to each state as follows. ⎧ 0, idle state ⎪ ⎪ ⎨ 1, loading and unloading state Ai = (9.2) ⎪ 2, processing state ⎪ ⎩ 3, demand state where the values 0 and 3 indicate that the CNC is in a non-working state; according to the goal of improving production efficiency, we naturally hope that the time of these two states is as short as possible. Working principle of RGV. Assuming that the conveyor belt can deliver the material to the feeding place in time, that is, there is no waiting time for the conveyor belt. Then in the whole course of working, the RGV is always located on the straight track, there are four states: waiting state, moving state, loading and unloading state and cleaning state. The waiting state indicates the state that RGV is resting on the track with no work; the moving state indicates the state when RGV is driven from one position to another; the loading and unloading state indicates the state in which RGV is feeding the CNC; and the cleaning state indicates the state when RGV is cleaning the material and placing it to the conveyor. RGV can operate in strict accordance with the operating instructions [5], set the state of the RGV as B and assign values to each state as follows. ⎧ 0, waiting state ⎪ ⎪ ⎨ 1, moving state B= ⎪ 2, loading and unloading state ⎪ ⎩ 3, cleaning state
(9.3)
68
M. Li et al.
where 0, 2, 3 states mean that RGV can only be in a fixed position on the track, at this time it can only be linked with the adjacent CNC of the position. Let j be where the RGV is, then ⎧ 1, i = 1 or 2 ⎪ ⎪ ⎨ 2, i = 3 or 4 j= ⎪ ... ... ⎪ ⎩ 10, i = 19 or 20
(9.4)
When the RGV is in 1 state, let its motion state be (j1 , j2 ), indicating that RGV is moving from position j1 to position j2 at this time. When the material is in the processing state, the corresponding CNC is in the working state. The goal of the planning is to improve the processing efficiency, that is, the more materials processed in the same time. Obviously, the shorter the non-working time of the CNC, the higher the processing efficiency. Determination of the objective function. Considering the shortest time for the CNC to be in a non-working state in one shift. For a complete machining process, the machining system starts at 0 s, RGV executes instructions from the initial position and at 28,800 s the machining system stops. For a given processing time, let n pairs of CNC be turned on (n = 1, 2, … 10) Define the total waiting time for CNC as follow. HI =
2n
h ik
(9.5)
i=1 k=1
where H I is the total waiting time of CNC in the Ith case, hik is the waiting time in the k operation of the CNC numbered i, and the material number is recorded according to the order in which the raw material is grabbed. By formula (9.5), in order to minimize the total waiting time of the CNC, the objective function is defined as the following optimization function. min HI
(9.6)
Initial scheme planning. Step 1: The initial case is considered firstly. The initial position of the RGV is position 1, and the CNC are all idle. The initial instruction is to feed those 2n CNC in certain order. Step 2: When the RGV is in a certain position, loading for the two adjacent CNC in that position does not need extra moving time. So when RGV is loading for one CNC at a certain location, the next loading target gives priority to another CNC at that location. Step 3: RGV first feeds the two CNC at position 1, then there are (n – 1)! possible initial feeding orders. Thus, the equivalent transformation of the objective function is as follows.
9 Research on the Optimum Ratio of Intelligent Rail …
min HI 1 , HI 2 , . . . , HI (n−1)!
69
(9.7)
Optimization model of scheduling scheme. Step 1: Taking the next move of RGV into account after finishing the feeding of 2n CNC for any initial plan. Although the cleaning step is added in the subsequent processing, the complete round of loading and unloading for 2n CNC will be given priority in the actual processing. So the dynamic planning of 2n consecutive complete operations is considered for RGV. Step 2: Since RGV is in a certain position, loading and unloading for the two CNC in that position do not need extra moving time, so it is considered that the next loading and unloading target of the RGV should give priority to the other CNC directly opposite the current one. That is, RGV will give priority to the two CNC at the same position in turn. Therefore, for the next 2n jobs, there are n2n types of scheduling schemes. Step 3: Assume that RGV has memory storage function and can record the time of last interaction with each CNC. For any scheduling scheme, each operation of RGV will result in a waiting time h im for the CNC numbered i, where m represents the mth of the 2n consecutive jobs. For this scheduling scheme, the total CNC waiting time is as follows. hi =
2n 2n
h im
(9.8)
m=1 i=1
The l here represents the lth 2n consecutive job in the whole process. Step 4: Find out the total waiting time under each scheduling scheme; there are n2n results. Take the scheme corresponding to the result with minimum total waiting time, that is, minimize the formula (9.8). The corresponding objective function are as follows. min h l =
2n
h il
(9.9)
i=1
Step 5: Choose the first step of the 2n consecutive jobs chosen above for RGV to execute. Then all CNC in this step have a total waiting time as follows. hh l =
2n
h il
(9.10)
i=1
Take the end of the step chosen above as the new beginning, repeat step 3 and step 4 and perform a number of 2n consecutive jobs until the end of a shift. For the vth initial case (v ∈ {1, 2, . . . , (n − 1)!}), the total waiting time for CNC resulting from the corresponding complete scheduling scheme is as follows.
70
M. Li et al.
HI v =
hh l
(9.11)
l=1
Step 6: Synthesize step 1 to step 5 and minimize the formula (9.11); the complete optimization model is as follows. min{HI v } ⎧ HI v = hh l ⎪ ⎪ ⎪ l=1 ⎪ ⎪ 2n ⎨ h il s.t hh l = i=1 ⎪ ⎪ ⎪ 2n 2n ⎪ ⎪ ⎩ min h l = h im
(9.12)
m=1 i=1
9.3.5 Determination of Optimum Ratio of CNC On the basis of practical experience, when the ratio of RGV and CNC is appropriate, they will not have to wait long for each other. So we consider choosing the shortest sum of total waiting time between RGV and CNC as the optimization goal. The optimal path planning model shows that when the number of the open CNC is 2n, the total waiting time H In of the CNC is also determined after the optimal path is determined. Let the corresponding number of finished materials be N In . To RGV, the assumption above shows that RGV won’t wait and not going anywhere after a cleaning task but immediately make a judgment and starts moving. Assume that the RGV works well, it moves strictly according to the instructions. When the RGV moves in front of the selected CNC, but the processing of this CNC is not completed, the waiting time of RGV will generate. Introduce the 0–1 variable Dik as follows. Dik =
1, the k-th feeding occur in the i-th CNC 0, else
(9.13)
During the kth feeding, the RGV arrives at the selected CNC at time t 1k , and the CNC completes the current material processing at time t 2k , introduce the 0–1 variable δ k as follows. δk =
1, t2k − t1k > 0 0, t2k − t1k ≤ 0
(9.14)
Let Q I n be the waiting time of RGV in a shift under the Ith case with 2n open CNC. Then
9 Research on the Optimum Ratio of Intelligent Rail …
QIn =
NI k 2n
Dik δk (t2k − t1k )
71
(9.15)
k=1 i=1
In summary, the objective function is as follows. min{HI n + Q I n }, n ∈ {2, 3, . . . , 10} n
(9.16)
9.4 Solution and Result 9.4.1 Solution of the Optimal Path Planning Model From the above, the set of processing time can be represented as follows. A = {T |T ∈ [10, 1200] and T |10}
(9.17)
For each case I that the processing time t ∈ A, consider the number of CNC from 1 to 10, respectively, to different scheduling scheme using traversal algorithm in MATLAB to solve. However, considering the amount of calculation and time, “five step by step” is used in the optimization model, which means in the initial state we get 2n steps as a process, then search the optimal path of five operations. After getting the processing scheme for the minimum total waiting time, compared the sum of the waiting time of RGV and each CNC corresponding to the optimal scheduling scheme for the different value of CNC number, and the optimal CNC number of the value of each machining time and the corresponding CNC working efficiency are finally obtained. Suppose that under the case I , the ideal number of finished materials is as follows.
T T − 2m I s (9.18) N I = 2m I T +s where [·] represents round down. Suppose that under the case I , the actual number of finished materials corresponding to n pairs of CNC is N I n , then there is a corresponding rate of finished materials. αI k =
NI k × 100% NI
and the maximum number of materials processed.
(9.19)
72
M. Li et al.
9.4.2 The Solution of Optimal CNC Number Ratio After the actual calculation, the sum of the processing time of RGV and each CNC is taken as the judgment index to balance the quantity of finished materials, the working efficiency of α I k and the quantity cost of CNC, so as to ensure the objectivity of the model and the rationality of the results. The result using traversal algorithm in MATLAB is shown in Fig. 9.2. According to the calculation results, when the processing time T ∈ [10 s, 210 s], the optimal number of CNC pairs is 1. When the processing time T ∈ [220 s, 360 s], the optimal number of CNC pairs is 2. When the processing time T ∈ [370 s, 510 s], the optimal number of CNC pairs is 3. When the processing time T ∈ [520 s, 650 s], the optimal number of CNC pairs is 4. When the processing time T ∈ [660 s, 800 s],
Fig. 9.1 Schematic diagram of intelligent processing system ( adapted from the “2018 Higher Education Cup National Mathematical Modeling Competition for College Students” problem B: intelligent RGV dynamic scheduling strategy)
Fig. 9.2 The optimum CNC quantity, working efficiency and total material number corresponding to the processing time of CNC single part
9 Research on the Optimum Ratio of Intelligent Rail …
73
the optimal number of CNC pairs is 5. When the processing time T ∈ [810 s, 940 s], the optimal quantity of CNC pairs is 6. When the processing time T ∈ [950 s, 1080 s], the optimal quantity of CNC pairs is 7. When the processing time T ∈ [1090 s, 1200 s], the optimal number of CNC pairs is 8.
9.5 Conclusions In this paper, the optimal path of RGV is planned by considering the future steps to determine the current step to approach the global optimal solution, which avoids complex global search and random generation with strong uncertainty, but still obtains the ideal result. When the optimal number of CNC is determined, this paper considers both processing efficiency and resource utilization, which is closely related to the actual production situation and has strong operability. According to the solution results, a general conclusion can be drawn: as the value of processing time increases, the optimal number of CNC increases. Meanwhile, from the calculation results, expect the results of poor preferability with the extreme value of the processing time T ∈ [10 s, 90 s], it can be found that the corresponding processing time amplitude of each optimal CNC number fluctuates within 110–140 s, and it is speculated that the corresponding relation has a certain periodicity. Whether this property is affected by the value of the time taken by RGV for CNC to load and unload materials, and the time taken by RGV to clean materials needs further research and explanation. This paper not only gives the optimal quantity ratio of CNC for different processing times, but also gives the corresponding scheduling scheme, which has a profound practical significance for production. However, when the processing time is shorter, namely less than 90 s, due to the large deviation in measuring the processing efficiency, the operability is not strong; hence, separate consideration should be given to the ratio problem in this processing time interval. In addition, the speed of the conveyor belt in actual production will also affect the scheduling scheme. In future studies, the influence of random faults can be considered, and the fault time and repair time can be simulated by Monte Carlo method. On this basis, the path planning of RGV can be carried out and the optimal ratio can be found. To sum up, this paper mainly uses dynamic programming to design the optimal scheduling scheme for the intelligent machining system and carries out the optimal ratio of the quantity of CNC and RGV according to different processing time. In this paper, the methods of “multi-step determination in the future of one step” and “screening based on processing efficiency and resource utilization” can be extended to more scheduling and matching problems so as to discuss the optimal grouping of CNC and RGV when the number of RGV increases.
74
M. Li et al.
References 1. Cai, X.D.: 2020 International Conference on Artificial Intelligence and Communication Technology (AICT 2020) 2. Zhang, G.Q., Zhang, Y.S.: Intelligent scheduling algorithm for linear reciprocating track automatic guide vehicle. Comput.Eng. 35(15), 176–178 (2009) 3. Shen, Y., Guo, B.: The network control system intelligent dynamic scheduling strategy with variable sampling period. J. Sichuan Univ. (Engineering Science edition), 42(1) (2020) 4. Zha, Z.Y., Li, J.X., Sheng, R.T., et al.: Application of intelligent translational track guide vehicle. Robot. Appl.5, 42–43 (2017) 5. Wu, G.H., Zhao, H.N., Zhu, X.C., Wu, Q.: 2019 6th International Conference on Machinery, Mechanics, Materials, and Computer Engineering (MMMCE 2019)
Chapter 10
An Apple Grading Method Based on Improved VGG16 Network Lin Kaiyan, Zhang Haoping, Liu Chang, Si Huiping, Wu Junhui, and Chen Jie
Abstract Apple grading is an important link of improving the commercial value of apples. This paper proposed an improved convolutional neural network M-VGG16 based on very deep convolutional networks (VGG16) to improve the accuracy and efficiency of apple grading. A convolutional layer of the Conv3, Conv4 and Conv5 was, respectively, deleted compared with VGG16, then a 1 * 1 convolution was connected in parallel with the three groups of convolutions. All convolutional layers were activated by ReLU function. A batch normalization layer was added between each convolutional and activation function layer. Under same dataset, the M-VGG16 network classification prediction accuracy rate is 96.3%, which is higher than AlexNet and VGG16. Compared with traditional machine learning methods, the classification effect of deep learning method has more significant advantages. Therefore, the apple grading method based on deep learning has the advantages of higher grading efficiency and stronger generalization ability, and it provides a new solution for apple classification.
10.1 Introduction Fruit is a major agricultural product in the world. As the largest fruit production and processing country in the world, China is known as the kingdom of fruit. Fruit processing has developed to be the third pillar agricultural industry in China with obvious advantages and competitiveness in agricultural product processing industry [1]. Improving the quality of fruits is a key factor for enhancing the competitiveness in the international market. Drawing on the experience of developed countries, fruit grading is one of important ways to increase production value. Apple classification and rating were usually handled with visual screening which was laborious and time consuming. With the development of machine learning technology, deep learning methods taking advantage of fruit image datasets to mine deep features and had brought new research ideas for solving traditional fruit grading L. Kaiyan · Z. Haoping · L. Chang · S. Huiping (B) · W. Junhui · C. Jie Modern Agricultural Science and Engineering Institute, Tongji University, Shanghai, China e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_10
75
76
L. Kaiyan et al.
problems [2–5]. Great progress had been made in the field of deep learning methods applied in fruits grading. Sa et al. [6] proposed a method for fruit detection utilizing deep convolutional neural networks and developed an object detector called faster region-based CNN (faster R-CNN) to detect and classify seven types of fruits. This method not only improves accuracy but also allows for rapid deployment. Muresan et al. [7] proposed a deep neural network that was capable of identifying 60 types of fruits. And the paper introduced a high-quality dataset named Fruit-360 that contained popular fruits for obtaining a good classifier. The network finally reached an accuracy of 96.3% on the test set. Kestur et al. [8] presented a MangoNet based on architecture of semantic segmentation. Compared with variant architectures of fully convolutional networks (FCN), the MangoNet had better detection robustness because it could eliminate the impact of illumination conditions, occlusion in the detection process. The network met the requirements for detection and counting in mangoes yield estimation. In conclusion, the reason why the deep network is faster and more accurate than traditional classification is to improve the architecture in new application scenarios. With convolutional networks becoming more popular in the field of computer vision, a number of works have been carried out to improve the original architecture. In this paper, we developed a supervised algorithm based on VGG16 in the bid to adapt to the particular fruit classification. The algorithm named modified VGG16 (MVGG16) could solve the problem of overfitting caused by less data or complicated model in the area of deep learning. Experimental results showed that the M-VGG16 need less time than Alexnet and VGG16 in the training period, but achieving a much higher prediction accuracy.
10.2 Material Preparation 10.2.1 Apple Grading Standards The apple grading standards showed in Table 10.1 were provided by the National Standard of the People’s Republic of China for Fresh Apples (GB/T10651-2008) [9]. The defects of apple mainly include stab, crush and insect fruit. According to the main features, apples are divided into such four grades as superior grade, first grade, Table 10.1 The standards of apple grading Features
1st grade
2nd grade
3rd grade
4th grade
Fruit diameter (mm)
≥100
≥85
≥70
0 and δ ∈ [0,1], Z = i=1 the RSD. H.264 Video Streams {R − 1} can be regarded as composed of multiple group of picture (GOP). Each GOP is composed of a certain number of I-frames, B-frames, and P-frames. I-frame is a complete image and does not need to refer to other frames when decoding. While B-frame and P-frame are predictive frames, which only store the motion vector rather than the complete image and need to refer to I-frame when decoding. In addition, the reference I-frame in the decoder is refreshed whenever a new I-frame arrives in the buffer, thus avoiding the spread of B-frame and P-frame decoding errors. Therefore, when encoding H.264 Video Streams, a group of GOP can be used as an independent input symbol unit, and LT codes encoding can be carried out combined with the RSD to obtain the encoded data package. In order to ensure the decoding success, there will be coding redundancy, which can be expressed by Eq. (23.4). ε=
K −k k
(23.4)
In Eq. (23.4), K represents the data packet encoded by the input symbol, k represents the input symbol, and ε represents the coding redundancy. As the packet loss rate increases, the loss of correlated coded data packets will lead to an increase in the probability of decoding failure. If the video data is encoded by the fixed coding redundancy, where a fixed code redundancy is applied for video data encoding, it is necessary to adopt a large data redundancy for the purpose of successful decoding of video data. This will result in waste of resources and reduction of efficiency when the packet loss rate is low. Therefore, when using LT codes to encode the video data, the coding redundancy should be determined according to the bandwidth situation, so as to realize the adaptation of coding redundancy and
194
R. Zhou et al.
network bandwidth and improve the coding flexibility and bandwidth utilization of LT codes.
23.2.2 Code Adaptive Coding Redundancy At present, bandwidth measurement methods can be divided into active measurement and passive measurement according to whether it is necessary to inject data packets into the network. The active measurement method requires continuous injection of additional data packets into the network, which will increase the link burden in network congestion. The detected data packets cannot be guaranteed to be reliable with UDP, so it is difficult to accurately measure the available bandwidth of the network. However, the video stream transmission has the characteristics of large amount of data and long duration, which enable us to estimate the bandwidth value by measuring the video stream directly. Reference [6] proposed a model to estimate the available bandwidth by calculating the GOP sending time of each group of video files. This algorithm can simply and quickly estimate the current network bandwidth by sending rate. This paper will combine this algorithm to construct LT code redundancy adaptive strategy. From the previous section, a GOP is a coding input unit, so the available bandwidth of wireless network can be estimated by the sending rate of a GOP. The size of a GOP is expressed as S j , the sending time is T j , j is the number of the current GOP, the bandwidth value Bj can be calculated by Eq. (23.5). Bj =
Sj Tj
(23.5)
In Eq. (23.5), a bandwidth can be estimated according to each GOP sent. Although the redundancy of fountain codes is not directly related to the bandwidth value, it shows a certain positive correlation with the packet loss rate. Therefore, it is necessary to establish the relationship between the bandwidth value and the packet loss rate. When the estimated bandwidth is greater than the video bit rate, it is considered that the current bandwidth meets the requirements of video stream transmission, and the packet loss probability is small. When the estimated bandwidth value is not greater than the video stream bit rate, it can be considered that the current bandwidth quality is poor and cannot satisfy the video stream transmission, and the probability of packet loss increases. Therefore, the relationship between coding redundancy ε and available bandwidth of wireless networks can be expressed by Eq. (23.6). ε=
a · (V − B), V − B > 0 0, V −B 0)
(24.1)
In Eq. (24.1), the time series Ui (t) = {u 1 , u 2 , ..., u t } and V j (t) = {v1 , v2 , ..., vt } were the sensing data acquired by the IoT terminals i and j during the time t and the Euclidean distance between the sequence data was d(u t , vt ), the dynamic time warping distance of the sequence was dDTW (Ui (t), V j (t)) = r (Ui (t), V j (t)).
24 Data Correction-Based Data Fusion Method for Internet …
201
The DTW has a high time complexity and a long-running time, so a sliding window of length t was set to analyze multiple terminal time data sequences [9]. The distance di, j between the terminal data sequences was expressed as Eq. (24.2): di, j (t) =
dDTW (Ui (t), V j (t)) i = j 0 i= j
(24.2)
where Ui (t) and V j (t) were the time data sequence of the terminal i and the terminal j within the window time t. Each terminal’s data in the divided data acquisition area should be positively correlated in time and space. If there was data with a low degree of correlation, it should be excluded. Median Absolute Deviation is a method for detecting outliers, and it was used to detect low-correlation terminals with large Spatial-temporal distances in this method. The median operation is set as med( ). This method calculated the median of the Spatial-temporal distance, then calculated the absolute difference with each distance, and finally took the median of all absolute values as MAD. It was expressed as follows: MAD = med di, j − med(di, j ) ( j= 1, 2,...,n)
(24.3)
Q = med(di, j ) + 2MAD
(24.4)
The threshold Q was set as the sum of the distance’s median and two times MAD. The acquisition terminal that exceeded the threshold was deemed low relevance and did not participate in the subsequent fusion calculations. The method improved the accuracy by fusing all associated terminal sensor data. The value of the distance di, j would appear to be 0, which was not suitable for weight assignment. Therefore, the exponential function was used to quantify the Spatial-temporal correlation degree si, j between each terminal data. The Z-score standardization method was used to process the Spatial-temporal distance di, j to conform to the normal distribution [10], and the exponential function was expressed as Eq. (24.5). si, j = f (di, j ) = e−di, j /2 2
(24.5)
24.2.2 Data Weighted Fusion Each acquisition terminal improved the accuracy of its sensing data by fusing other related terminals’ sensing data. According to the Spatial-temporal correlation si, j between these terminals, weights were assigned to perform weighted data fusion.
202
R. Lei et al.
The fusion weights were allocated using the Spatial-temporal correlation obtained by Eq. (24.5), where the higher the correlation between the terminal sensor data, the greater the weight assigned. si, j ωi, j = n
k=1 si,k
(24.6)
In the equation, i, j and k were the labels corresponding to the terminal, and the weight ωi, j should satisfy: ⎧ N ⎨ ωi, j =1 j=1 ⎩ 0 < ωi, j ≤ 1
(24.7)
According to the calculated weights, weighted data fusion was performed on the sensor data of each acquisition terminal at time t, and the result is shown as: xi∗ (t) =
N j=1
x j (t)ωi, j
(24.8)
Data x j (t) was the actual measured value of each terminal at time t, and xi∗ (t) was the fusion result of terminal i at time t. The fusion results were used as the final measured value to correct the original data with low accuracy.
24.3 Simulation Results and Analysis Through two simulation experiments, the effective and reliability of the DF-BST method were verified. The first experiment used six sensor mathematical models to generate random datasets and used their random errors as simulated environmental interference. The second experiment used the real datasets in [11] for data fusion and compared with the experimental results. These simulation experiments were performed in MATLAB 2017a, and used mean absolute error (MAE) and root mean square error (RMSE) to measure the accuracy of the method. The reliability of the method was verified by setting up six mathematical models of sensors belonging to different terminals and completing experiments. The sliding window size was set to 10. The simulation data was Si , and x was the real value of 20. The equation was shown as follows:
24 Data Correction-Based Data Fusion Method for Internet …
⎧ S1 ⎪ ⎪ ⎪ ⎪ ⎪ S2 ⎪ ⎪ ⎪ ⎨ S3 ⎪ S4 ⎪ ⎪ ⎪ ⎪ ⎪ S5 ⎪ ⎪ ⎩ S6
203
= x + n1 = x + n2 = x + n3 = x + n4
(24.9)
= xe0.0007x + n 5 = x + sin(0.1x) + n 6
Random noise n i (i = 1, 2, ..., 6) obeyed a normal distribution with a mean value of 0 and a variance of 0.1. Gross errors were added to S5 and S6 to simulate external interference. Based on the simulation datasets, the DF-BST method, the mean method and the method in [11] were used for simulation experiments. The results were shown in Fig. 24.1 and Table 24.1. Figure 24.1 shows that the result of the DF-BST method is closer to the real value, and the experimental result is better than other methods. The reason is that the two sensor models in this experiment increased gross errors, and the DF-BST method used the temporal and spatial correlation between the data to calculate, which can effectively reduce the errors caused by external interference.
Fig. 24.1 Results of data fusion
Table 24.1 Comparison of data fusion results based on datasets fusion results
RSME
MAE
The mean method
0.1145
0.1084
DF-BTW
0.0590
0.0480
DF-BST
0.0448
0.0362
204 Table 24.2 Comparison of data fusion results based on real datasets
R. Lei et al. Fusion results
Errors
The mean method
54.3967
0.02609
DF-BTW
52.7440
0.00483
DF-BST
53.0792
0.00179
As shown in Table 24.1, the DF-BST method is better than other methods in indicators such as MAE and RSME. Therefore, according to the fusion results and various indicators, it can be concluded that the DF-BST method is effective and has better performance. The DF-BST method was used for calculation based on the datasets in [11]. The true value of the data was 53, and the comparison results were shown in Table 24.2. As shown in Table 24.2, the error of the DF-BST method is lower. Because the DF-BTW method only considers the relevance of the sensor data in time, and the DF-BST method considers the relevance of the sensor data in time and space, the result is more accurate.
24.4 Conclusion This article proposed a data correction method-based data fusion for IoT terminals, which used the temporal and spatial correlation between IoT sensor data to improve data accuracy. Simulation results show that this method can effectively improve data accuracy. This method uses the Spatial-temporal correlation between similar data for data fusion, but the correlation between different sensor data types has not yet been considered. It will continue to be studied and improved as a future research direction.
References 1. Feiyue, W., Jun, Z., Xiao, W., et al.: Industrial intelligent networking: basic concepts, key technologies and core applications. IEEE/CAA J. Automatica Sinica 44(9), 1606–1617 (2018). (in Chinese) 2. Lee, Y.T., Hsiao, W.H., Lin, Y.S., et al.: Privacy-preserving data analytics in cloud-based smart home with community hierarchy. IEEE Trans. Consum. Electron. 63(2), 200–207 (2017) 3. Liqin, T., Chuan, L., Qi, Z., et al.: Reliability design and optimization analysis of IoT monitoring topology. J. Softw. 25(08), 1625–1639 (2014). (in Chinese) 4. Yue, Y., Zuo, X., Luo, X.: Biased estimation method of multi-sensor data fusion to improve measurement reliability. IEEE/CAA J. Automatica Sinica 40(09), 1843–1852 (2014). (in Chinese) 5. Yu, X., Fan, F., Zhou, L., et al.: Adaptive predictive weighted data fusion algorithm for wireless sensor networks. Chinese J. Sens. Actuators 30(05), 772–776 (2017). (in Chinese) 6. Wu, H., Gao, S., Peng, H., et al.: Adaptive fuzzy C-means clustering data fusion algorithm. Comput. Eng. Appl. 55(05), 26–35+82 (2019). (in Chinese)
24 Data Correction-Based Data Fusion Method for Internet …
205
7. Gressent, A., Malherbe, L., Colette, A., et al.: Data fusion for air quality mapping using low-cost sensor observations: feasibility andadded-value. Environ. Int. 143,105965 (2020). (in Chinese) 8. Li, Z., Zhang, F., Li, K., et al.: A multivariate time series index structure supporting DTW distance. J. Softw. 25(03), 560–575 (2014). (in Chinese) 9. Feng, Y., Ding, C., Gao, X., et al.: Time series similarity based on moving average and piecewise linear regression. Comput. Sci. 45(S1), 110–113 (2018). (in Chinese) 10. Bi, J., Zhen, J., Wang, Y., et al.: Improved KNN indoor positioning method based on Gaussian function. Bull. Surveying Mapping (06), 9–12+35 (2017). (in Chinese) 11. Hu, W., Zhao, W.: Energy Internet data fusion method based on time window and adaptive weighting. Syst. Engi.-Theory Methodol. Appl. 25(05), 907–913 (2016)
Chapter 25
Research on Multi-Sensor Data Fusion Algorithm for Monitoring of Power Distribution Station Peng Wang, Min Xiang, and Rujie Lei
Abstract Aiming at the issue of data anomalies and excessive measurement errors caused by the limitation that monitoring nodes in power distribution stations are vulnerable to the environment, an improved data fusion algorithm based on closeness concept is proposed in this research. First, the Grubbs criterion is used to validate the validity of the data, the invalid data is removed to form a set of valid values. Then, in order to reduce the influence of measurement error on the fusion accuracy, the support degree function is introduced to improve the method of target value selection and valid value processing in the closeness-based fusion algorithm. The simulation test results show that the algorithm can effectively reduce the mean square error of the fusion result and improve the fusion accuracy.
25.1 Introduction With the construction of smart grid, ubiquitous power Internet of Things projects and the development of Internet of Things technology, advanced smart sensor technology and data processing technology have been applied to monitor the environment and equipment in the distribution station, and to a certain extent, the efficiency and reliability of the monitoring process have been improved [1–4]. Hong-Xing et al. [5] designed a wireless temperature measurement system for the bus-bar temperature of the switch-gear to realize the online monitoring of the bus-bar temperature of the switch-gear in response to the issue of low safety of temperature measurement of the switch-gear bus-bar in the distribution station; [6] used multiple sensors and a data fusion method based on closeness to fuse the monitoring data of power equipment, this algorithm improved the data accuracy to a certain extent, but does not consider the measurement error caused by the distribution station environment, so the fusion accuracy will be reduced. P. Wang (B) · M. Xiang · R. Lei Key Laboratory of Industrial Internet of Things and Networked Control, Chongqing University of Posts and Telecommunications, Chongqing 400065, China e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_25
207
208
P. Wang et al.
The algorithm proposed in this research introduced the concept of comprehensive support degree and improved the data fusion algorithm based on closeness to suit the special environment of power distribution station. This algorithm uses Grubbs criterion to check the validity of data to reduce the impact of invalid data on the monitoring results caused by the environmental interference of the distribution station; The comprehensive support degree is introduced to improve the fusion algorithm based on the closeness degree, so as to reduce the interference of measurement error and improve the accuracy of the fusion result.
25.2 Data Fusion Algorithm 25.2.1 Fusion Algorithm Base on Closeness Assuming that the original measurement data obtained from the n nodes monitoring the temperature of the power distribution room is xi (i = 1, 2, . . . n), the closeness between the measurement data and the target value can be obtained through the closeness function based on the distance. The target value should theoretically be the actual value, but the actual value is unknown in practical applications, so the algorithm usually uses the average value of the measured data [7]. Assuming that the closeness between the original measurement value xi and the target value x is Ri (0 ≤ Ri ≤ 1), the fusion algorithm bad on closeness is as follows: ⎧ ⎨ Ri = 1 1 + di ⎩ di = |xi − x|
(25.1)
In Eq. (25.1), Ri is the closeness; di represents the distance between each original measurement value and the target value. Then use Eq. (25.2) to assign weight ωi to each original measurement value xi , and finally obtain the fusion result Y through Eq. (25.3). ωi =
Ri n Ri
(25.2)
i=1
Y =
n i=1
xi ωi
(25.3)
25 Research on Multi-Sensor Data Fusion Algorithm for Monitoring …
209
25.2.2 Algorithm Improvement Data preprocessing. This research used Grubbs criterion for data validity judgment. Assuming that the original data collected in the distribution room temperature was xi (i = 1, 2, . . . , n), the Grubbs statistic gi of each data could be calculated by Eq. (25.4), among them, xm and σ are the median and standard deviation of the data set. gi =
|xi − xm | σ
(25.4)
Then according to the number of data n and Grubbs significance level a can determine the critical threshold gt (n, a). If gi ≥ gt , the data is considered invalid and should be eliminated. After preprocessing, the remaining data constituted the effective value set yi (i = 1, 2, . . . , n), m ≤ n. Improvement based on comprehensive support degree. In this research, the support degree function sup(a, b) was used to express the degree of support of b to a. The support degree function proposed by Yager must meet the following three conditions [8]: (1) (2) (3)
sup(a, b) ∈ [0, 1];[0, 1]; sup(a, b) = sup(b, a); if |a − b| < |x − y|, sup(a, b) > sup(x, y).
Considered the high complexity of the support degree function proposed by Yager, in order to reduce the complexity, this research used the support degree function proposed in [9] to describe: ri j = sup(yi , y j ) =
K 1 + β(yi − y j )4
(25.5)
where yi and y j represent the measure values of node i and node j; K represents the magnitude of the support degree function; β ≥ 0, is the attenuation factor; Then the degree of the support ri j and support degree matrix D could be obtained: ⎡
r11 ⎢ r21 ⎢ D=⎢ . ⎣ ..
r12 r22 .. .
··· ··· .. .
r1m r2m .. .
⎤ ⎥ ⎥ ⎥ ⎦
(25.6)
rm1 rm2 · · · rmm
The comprehensive support degree of node i is shown in Eq. (25.7): m
si = ri j j=1
(25.7)
210
P. Wang et al.
By Eq. (25.8), we could find the value x h with support degree of smax and the value xl with support degree of smin in the effective value set, among them, smax and smin represent the highest and lowest comprehensive support degree, respectively, and the lowest comprehensive support degree indicates that the data has the largest measurement error in the effective value set, and the opposite indicates the smallest measurement error. In order to reduce the interference of the distribution station environment on the fusion result, the method of selecting the target value and the effective value processing in this research was improved. First, selected the value x h with the comprehensive support degree smax as the target value to realize the selection of the target value; then, in the effective value set, the value xl whose comprehensive support degree was smin replaced with the value x h whose comprehensive support degree was smax , and the processed data z i could be obtained. ⎧ ⎪ ⎨ smax = max{s1 , s2 , . . . , sm } smin = min{s1 , s2 , . . . , sm } (25.8) ⎪ ⎩ z i (i = 1, 2, . . . , m), m ≤ n Fusion result calculation. Through Eqs. (25.5)–(25.8), the target value x h and the processed data z i could be obtained. Replaced the original measurement data xi in Eq. (25.1) with z i , and replaced the original data average value x with x h , Eq. (25.9) could be obtained. ⎧ ⎨ Ri = 1 1 + di (25.9) ⎩ di = |z i − x h | Then, replace the original measurement data xi in Eqs. (25.2) and (25.3) with z i , and the weight ωi and the fusion result Y of the processed data z i could be obtained, respectively: ωi =
Ri m Ri
(25.10)
i=1
Y =
m i=1
z i ωi
(25.11)
25 Research on Multi-Sensor Data Fusion Algorithm for Monitoring …
211
25.3 Simulation Test and Analysis In order to validate the feasibility of the algorithm, this research used the experimental data of [10] for calculation and analysis, the specific perception data was xi (i = 1, 2, . . . , 10), which were: 1.00, 0.99, 0.98, 0.97, 0.5, 0.65, 1.01, 1.02, 1.03, 1.5, and the actual value of this parameter was 1.00. First, this research used Eq. (25.4) to judge the data validity. Among them, the Grubbs statistics of data with node numbers 5, 6, and 10 were 2.942, 2.706, and 2.039, respectively, which were all greater than the Grubbs critical threshold 2.036, so the data of nodes 5, 6, and 10 would be eliminated. Then used Eqs. (25.5)–(25.7) to calculate the comprehensive support degree of the remaining data. From Eq. (25.8), we could get that the comprehensive support degree of node 4 and node 7 was the lowest, which was 6.99998, and the comprehensive support degree of node 1 data was the highest at 7.00. According to the algorithm, the data of node 4 and node 7 need to be replaced with the data of node 1, and substituted the data of node 1 as the target value into Eqs. (25.9) and (25.10) to calculate the weight of each data, respectively: 0.14407, 0.14265, 0.14104, 0.14407, 0.14265, 0.14104, and 0.14407. Finally, the fusion result could be calculated by using formula (25.11): Y = 1.00 × 0.14407 + 0.99 × 0.14265 + 0.98 × 0.14104 + 1.00 × 0.14407 + 1.01 × 0.14265 + 1.02 × 0.14104 + 1.00 × 0.14407 = 0.99958 Among them, the result obtained by using the closeness-based fusion algorithm was 0.9679, and the result of [10] was 0.9994. It could be seen from the above results that the result of the algorithm in this research was closer to the actual value of 1.00, so it was feasible to use the algorithm in this research for fusion. In order to further verify the effect of the algorithm, this research used the data in [6] for fusion. The actual value was 24 °C. Copied some data of this literature to get 12 sets of data waiting to be merged, and introduced 1 randomly generated abnormal data in groups 5~8, and 2 randomly generated abnormal data in groups 9~12 to simulate the invalid data generated by the environmental interference of the distribution station. The specific data is shown in Table 25.1. The accuracy of the fusion algorithm generally uses the mean square error as the evaluation index, through the fusion calculation of the data in Table 25.1, compared the fusion results and mean square error of the improved fusion algorithm based on the closeness degree in this research, the closeness-based fusion algorithm in [6], and the arithmetic average method. The comparison result is shown in Fig. 25.1. As shown in Fig. 25.1a, the fusion result of the improved closeness-based fusion algorithm is closer to the actual value than the closeness-based fusion algorithm and arithmetic average method, and less fluctuation. The mean square error comparison is shown in Fig. 25.1b. The improved closeness fusion algorithm has the smallest mean square error, followed by the closeness-based fusion algorithm, while the arithmetic average method has a large and unstable mean square error. The above comparison
212
P. Wang et al.
Table 25.1 Data to be fused Group number
Node 1
Node 2
Node 3
Node 4
Node 5
Node 6
Node 7
Node 8
1
23.7
23.6
23.7
23.9
23.4
24.1
23.8
23.9
2
23.5
23.8
23.7
23.9
23.5
24.2
23.9
23.8
3
23.7
23.7
23.7
23.2
23.4
24.1
23.9
23.8
4
23.5
23.8
23.2
23.7
23.7
24.1
23.8
24.2
5
23.5
23.6
23.7
23.9
23.4
24.1
23.8
23.9
6
22.7
23.8
23.7
23.9
23.5
24.2
23.9
23.8
7
25.2
23.7
23.2
23.2
22.3
24.1
23.9
23.8
8
23.7
23.8
23.7
23.7
23.7
24.1
23.8
22.1
9
23.5
23.6
22.3
22.3
23.4
24.1
23.8
23.9
10
25.3
25.3
23.9
23.9
23.5
24.2
23.9
23.8
11
23.7
23.7
23.7
22.1
23.4
24.1
22.1
23.8
12
22.4
22.4
23.2
23.7
23.7
24.1
23.8
24.2
Fig. 25.1 Comparison of fusion results and mean square error of different algorithms
results show that the algorithm in this research has obvious advantage infusion accuracy, the reason is that the data preprocessing mechanism of the algorithm in this research reduces the influence of invalid data on the fusion result. On the other hand, the improvement of the closeness fusion algorithm through the support degree function further reduces the influence of measurement errors on the fusion result and improves the accuracy of the fusion.
25 Research on Multi-Sensor Data Fusion Algorithm for Monitoring …
213
25.4 Conclusion In order to solve the issue that the monitoring node of the distribution station is susceptible to its environmental interference, which causes data anomalies and measurement deviations, resulting in low fusion accuracy, an improved data fusion algorithm based on closeness was proposed. Firstly, the Grubbs criterion is used to identify and eliminate the invalid data generated in the monitoring process, which improves the algorithm’s ability to suppress outliers. Secondly, the introduction of comprehensive support degree improves the fusion algorithm based on closeness degree, which effectively reduces the influence of measurement error on the fusion result. The simulation test results show that the fusion result of the algorithm has a low mean square error.
References 1. Yi, W., Qi-Xin, C., Ning, Z., Cheng, F., et al: Fusion of the 5G communication and the ubiquitous electric internet of things: application analysis and research prospects. Power Syst. Technol. 43(05), 1575–1585 (2019) 2. Gang, L., Bo, Z., Wen-Qing, Z., Yun-Peng, L., Shu-Guo, G.: Data science issues in power equipment condition assessment: challenges and prospects. Autom. Electric Power Syst. 42(21), 10–20+177 (2018) 3. Ting, Y., Feng, Z., Ying-Jie, Z., Hai-Bo, P.: Interpretation and research prospects of ubiquitous power internet of things. Autom. Electric Power Syst. 43(13), 9–20+53 (2019) 4. Dong-Eun, K., Hyun-Jae, L., Jin-Geun, S., Jae-Don, P.: Modified expert inference method of power substation monitoring system based on expansion of multi-sensor utilization for fire discrimination. J. Electr. Eng. Technol. 14(3), 1385–1393 (2019) 5. Hong-Xing, F., Feng-Liang, G., Yu-Yao, X., et al: Wireless acquisition system for busbar temperature of switch cabinet based on STM32. Instrum. Tech. Sens. 4, 77–80 (2020) 6. Shang-Cong, Z.: Research on Data Fusion Algorithm in Wireless Sensor Networks for Power Equipment Temperature Monitoring. Shanghai Dian Ji University, Shanghai (2018) 7. Yuan, Z., Jiu-Gen, Z., Jia-Le, L., Xin, C.: Fuzzy nearness recommendation algorithm based on heterogeneous information network. Comput. Eng. Design 41(02), 367–372 (2019) 8. Ronald, R.Y.: The power average operator for information fusion. In: 13th International Conference on Information Processing and Management of Uncertainty, Dortmund (2010) 9. Wen-Guang, Y., Yun-Jie, W., Jian-Min, W.: Improvement and Development of the New Model of Grey Correlation Analysis. J. Zhengzhou Univ. (Nat. Sci. Edn.) 49(02) 25–30 (2017) 10. Jian-Shu, L., Ren-Hou, L. Hong, C.: Multi-sensor data fusion based on correlation function and least square. Control Decis. 6, 714–716+720 (2006)
Chapter 26
Analysis of QR Reordering Algorithm Based on Feedback Technology Optimization You Xinhua, Long wenjia, and Dai Yu
Abstract Feedback technology is an important tool to understand user search intentions, achieve personalized ranking, improve search engine ranking results, and increase user satisfaction with search engine results. Using feedback information to optimize the index of sorting performance and using it for resorting algorithms is a current research hotspot. This paper deeply studies the characteristics of users’ feedback behavior and proposes a new evaluation indicator Quality of a Ranked list (QR). Then, on the basis of the existing classification evaluation method, the re-ranking process is improved and proposed a new and stable reordering algorithm which is introduced. The reordering algorithm proposed in this paper avoids the unreliability and noise problems of feedback documents.
26.1 Introduction A typical user search method is that users use search engines as an information service form according to their needs to help users quickly and accurately find what they need in the ocean of information. The search purpose is organized into several keywords as input and sent to the search engine. The search engine searches according to the keywords and returns the search results to the user [1–4]. However, a query result often contains thousands of uneven and different contents. Often only a few of these content are in line with the user’s retrieval purpose, and the accuracy of the query result set is low. Because users cannot organize their search intent into high-quality keywords, the search results are unsatisfactory. Therefore, search engines need some additional technologies to assist in mining users’ query intentions and improve search results [5–7]. The use of feedback technology can be closer to the user’s search intention, so that users can get the desired search results and get personalized search. The classic information retrieval model of personalized search usually consists of four elements: the logical representation of a large document collection; the logical representation of user information needs; the background description of the model; Y. Xinhua · L. wenjia · D. Yu (B) Zhixing College of Hubei University, Wuhan, China e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_26
215
216
Y. Xinhua et al.
and the sorting algorithm [8–11]. For any user information requirement, the sorting algorithm calculates the relevance weight of each document [12]. The main function of the relevant feedback technology is to provide a new sorting algorithm, in the case of adding a known condition (user feedback information collection), recalculating the relevance weight of the query document, thereby changing the original sorting result, improve the quality of retrieval. After getting user feedback, how to evaluate the results is an important step in related feedback technology. In the field of information retrieval, the two most classic evaluation indicators: precision and recall [13–16]. Accuracy is a measure of the ratio of found-related documents in all returned result sets, and the recall rate is a measure of the ratio of found-related documents in all related documents. However, the accuracy and the recall rate only measure the quality of the result set and do not consider the quality of the result set sorting. Literature proposed an evaluation index to measure the quality of ranking: Discounted cumulative gain (DCG) index [17–20]. Using this indicator, the order of sorting in a document collection can be evaluated. In any system that uses user feedback, because the user gives a relevance judgment on certain documents, the system naturally prioritizes the ranking of related feedback documents. The result will definitely lead to an improvement in accuracy or recall rate, that is, “sort “Effect” improved. However, because the ranking of the documents that the user has not yet viewed may not be improved, that is, the “feedback effect” has not been improved, then such feedback results do not bring any benefit to the user. Therefore, when evaluating a feedback system, it is necessary to distinguish between “ranking effect” and “feedback effect.” Based on the existing “ranking method” and “feedback effect,” this paper proposes a measurement indicator Quality of a Ranked list (QR). Then, on the basis of the existing classification evaluation method, the re-ranking process is improved and the A new stable reordering algorithm.
26.2 QR Index Reordering Algorithm The sorting method of documents directly affects the “feedback effect” and ultimately affects the retrieval results of documents. Usually precision, recall, and other derivative indicators, such as MAP (Mean Average Precision) [21, 22], r-p (RecallPrecision), etc., are used to measure the quality of a sorting result. Precision refers to the proportion of documents retrieved that are related documents; recall rate refers to the proportion of how many related documents are retrieved correctly. The calculation method of MAP is: the MAP of a single query subject is the average of the similarity of each related document. The above indicators are based on the set and do not consider the impact of the ranking results. How to measure the quality of the ranking results, using the DCG indicator. The calculation formula of this indicator is Formula 26.1:
26 Analysis of QR Reordering Algorithm Based on Feedback …
DCGi =
if(i = 1) G1, DCGi−1 + G i / logb i, otherwise
217
(26.1)
DCGi represents the DCG value of the i-th document in the query result set, represents the similarity weight of the i-th result, and b is a parameter. It can be seen from Formula 26.1 that the importance of the result scores with the lower sequence number decays according to the logarithm of the sequence number. The biggest advantage of the DCG indicator is that it has a variety of relevance levels, that is, the relevance of documents is not only represented by 0,1, but is uniformly represented in an indicator. The higher the ranking of the document, the greater the contribution to the DCG value. After the relevance of each result is manually marked, the DCG index can be used to calculate the quality of the current ranking result. However, the DCG index can only be applied to the experimental process. In the real retrieval process, DCG cannot be used to evaluate the quality of a ranking result. This paper proposes a new indicator QR, which is proportional to the DCG indicator and can be calculated without any manual intervention. In an actual retrieval process, the QR indicator is a real and calculable indicator. The previous reordering algorithms mainly focused on the relevance of a candidate document to user interests. Score each candidate document according to certain indicators and sort all candidate documents according to the score. The focus is to examine the quality of a document and examine the similarity between a single document and the user’s intention. This article solves the reordering problem from another perspective: enlarge the granularity of the object to be examined and examine the characteristics that a document sequence (a ranking result) should have when the overall relevance of it is better. The measure of the quality of a ranking result can be calculated from the characteristics of the ranking result. A retrieval intention of a given user q and an initial document collection {d1 , d2 , . . . , dn }, let the ideal sorting result is d1 , . . . di , di+1 . . . , dn . In this ideal ranking, all documents are arranged in descending order of relevance to the user’s search intent; d1 , . . . di are related documents (that is d1 , . . . di , their relevance is greater than the user’s minimum expected threshold for retrieval results), and di+1 . . . , dn are irrelevant documents. By studying the ideal result set, it can be assumed that the ideal result set should have the following characteristics: Sort: Documents in {d1 , d2 , . . . , dn } are sorted in descending order of relevance. Since the similarity is only the result of guessing, this is the most basic and easiest condition to satisfy, and it is also an impossible condition. The QR measurement index proposed in this article only needs to complete a descending order at the end. Radius: The average distance from each document to the centroid is radius in d1 , . . . di . radius =
relnum 1 dist d j , dcenter relnum j=1
(26.2)
218
Y. Xinhua et al.
where radius represents the radius of the first i documents; relnum represents the number of related documents i; the dist function is a distance function, used to calculate the distance between the document d j and the centroid d center of the previous i documents. In the ideal result sequence, all related documents are ranked first, and they are more similar to each other (greater than the similarity to unrelated documents). Therefore, in a ranking result with better quality, if more related documents are ranked higher, the radius of the first i documents is also smaller. Document ratio: In the document collection di+1 . . . , dn , that is sorted lower, count how many documents may be similar documents. Calculated by Formula (26.3), where ratio represents the document ratio, d center represents the centroid of the similar document d1 , . . . di set, radius represents the radius of the similar document set, and irelset represents the irrelevant document set, that is di+1 . . . , dn . Assuming that d j is sufficiently similar to similar documents (that is, the distance between d j and the centroid d center of the similar document set is less than the radius of the similar document set), it is considered that d j may be a similar document that has been missed. The more similar documents that are missed (that is, the higher the proportion of documents), the worse the quality of the resulting sequence. Combine the above three features to construct QR indicator: QR(Dn ) =
β α + radius ratio
(26.3)
where Dn represents the sorting quality of a result sequence {d1 , d2 , . . . , dn }, radius represents the radius of related documents, ratio represents the possible proportion of related documents, and is an empirical adjustment parameter used to adjust the influence of radius and ratio on the final QR value. Usually n is 200, which means that the first 200 search results are reordered. Although for a query, search engines may return thousands of search results; on the one hand, users often only care about the sorting results of the first few pages; on the other hand, because the re-sorting process must be completed in real time; therefore, in a user feedback, in the system, the processed document collection is usually only the initial search results of the first few pages. In such a document collection, it can be seen that the value range of the radius and the value range of the ratio document ratio are quite different. In order to balance the influence of these two factors on the QR indicator, an empirical adjustment function log is added, and the QR calculation formula is: 1 1 + (26.4) QR(Dn ) = log radius ratio The QR indicator is directly proportional to the DCG indicator, which means that the QR indicator can measure the quality of a ranking result.
26 Analysis of QR Reordering Algorithm Based on Feedback …
219
26.3 Improved Reordering Algorithm Using QR Indicator The usual user feedback algorithms are mainly divided into two types: user feedback algorithms based on optimized queries and user feedback algorithms based on document comparison. There are three main reordering algorithms for user feedback: (1)
IDE-Regular algorithm [12]. A classic query expansion and weight adjustment algorithm in the vector space model (VSM). Ide-Regular: qm = α q + β
dj − γ
∀d j ∈Dr
(2)
dj
(26.5)
∀d j ∈Dn
IDE-Regular algorithm is a basic algorithm for user feedback based on optimized queries, where qm is new adjusted query, q is original query, Dr is collection of relevant user feedback documents, and Dn is the user’s negative feedback document collection, β = 0.75, γ = 0.25, α = 0.25. Linear comparison algorithm. Given n feedback documents { f 1 , f 2 , . . . , f n }, the relevance of candidate document d i is obtained by linear addition of n (wherein represents the similarity between document d i and feedback document f i ). Use it as a basic algorithm for user feedback based on document comparison. relevance of(di | f 1 , f 2 , . . . f n ) =
n
w f j ,di
(26.6)
j=1
(3)
QR weighting algorithm. This algorithm is proposed in combination with the QR indicator. For each feedback document, the quality of the current ranking result that is QRcurrent value when it is clicked is calculated; this is used to measure the reliability of different feedback documents. When the QR value is high, the predicted relevance of the document is high, and it should have a high influence on the reordering process, so the weight is relatively large. relevance of(di | f 1 , f 2 , . . . f n ) =
n
QRcurrent w f j ,di
(26.7)
j=1
On the basis of obtaining QR, the sorting algorithm is further improved. In most cases, the user’s click data is directly used as the relevant feedback information. However, a result document being clicked by the user does not mean that the result document must be a relevant document. There are many reasons that affect the user’s click behavior, and the quality of the current sorting result is one of the main reasons: when the current sorting result is of good quality, the user’s click document can well reflect the user’s needs, when the quality of the current ranking results is not as good as expected, the user will take the second place and click on some less relevant documents, that is, “pull generals from short ones.” This phenomenon is
220
Y. Xinhua et al.
Fig. 26.1 Reordering algorithm flow
called quality deviation, that is, the quality of the current ranking results affects the quality of the documents that users click on. Therefore, the processing of user click data should consider the quality of the current ranking results. The current ranking result is of better quality, that is, when the QR value is high, and the feedback document clicked by the current user may be more relevant; the feedback document should have a greater influence on the re-ranking process. Therefore, the core idea of the new reordering algorithm is to use the QR value to weight different feedback documents to reflect the different importance of each feedback document. The flow of the algorithm and its relationship with general search engines is shown in Fig. 26.1.
26.4 Conclusion Based on the in-depth study of user feedback behavior characteristics, in order to avoid the unreliability and noise problems of feedback documents, this paper introduces classification technology to improve the reordering process and improve the efficiency of the classification algorithm. After carefully evaluating and examining the existing evaluation methods for classification, a new stable evaluation method QR algorithm is proposed to improve the quality of the ranking results. On the basis of the QR indicator, the relevance of different feedback documents can be calculated by the quality of the current ranking results, which is the idea of the QR weighted ranking algorithm. In this way, the feedback documents with higher relevance play a more important role in the reordering process and ultimately effectively improve the quality of the ranking results. Acknowledgements This work described in this paper was fully supported by a grant from the Guiding Project of Science and Technology research plan of Hubei Provincial Department of Education (No: B2020329).
26 Analysis of QR Reordering Algorithm Based on Feedback …
221
References 1. Baeza-Yates, R., Ribeiro-Neto, B.: Modern Information Retrieval. Pearson Education (2014) 2. Ruthven, I.: Re-examining the potential effectiveness of interactive query expansion. SIGIR’09 Proceedings of the 32th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 213–220, Boston, USA (2012) 3. Dumais, S., Cutrell, E., Cadiz, J.J., Jancke, G., Sarin, R., Robbins, D.C.: Stuff I’ve Seen: A system for personal information retrieval andre-use. Proceedings of SIGIR (2016) 4. Morita, M., Shinoda, Y.: Information filtering based on user behavior analysis and best match text retrieval. ACM Conference on Research and Development in Information Retrieval (SIGIR) 49(2), 28–35 (2015) 5. Joachims, T., Granka, L., Pan, B.: Accurately interpreting clickthrough data as implicit fe. SIGIR 2005 51(1), 4–11 (2016) 6. Lundquist, C., Grossman, D.A., Frieder, O.: Improving relevance feedback in the vector space model. Proceedings of the CIKM Conference (CIKM97), LasVegas Nevada, USA (2010) 7. Rocchio, J.J.: Relevance feedback in information retrieval. The Smart Retrieval SystemExperiments in Automatic Document Processing. G.Salton, Prentice Hall Inc., Englewood Cliffs, NJ (2016) 8. Ide, E.: New experiments in relevance feedback, pp. 337–354. The Smart Retrieval System. G. Salton, Prentice Hall (2012) 9. Lee, J.H.: Combining multiple evidence from different relevance feedback methods. Inf. Process. Manag. 34(6), 681–691 (2017) 10. Yu, C.T., Buckley, C.: A generalized term dependence model in information retrieval. Inf. Technol. Res. Develop. 2(4), 129–154 (2016) 11. Ponte, J.M., Croft, W.B.: A language modeling approach to information retrieval. SIGIR’98 In 21st ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR’98) (2010) 12. Hiemstra, D., Kraajj, W.: Twenty-one at TREC-7:ad-hoc and cross-language track. Proceedings of Seventh Text Retrieval Conference (TREC-7) (2015) 13. Kurland, O., Lee, L., et al.: Better than the real thing: iterative pseudo-query processing using cluster-based language models. IGIR’05 (2005) 14. Kelly, D., Teevan, J.: Implicit Feedback for Inferring User Preference: A Bibliography. sigir forum (2003) 15. Baeza-Yates, R., Ribeiro-Neto, B.: Modern Information Retrieval. Pearson Education (2004) 16. Ruthven, I.: Re-examining the potential effectiveness of interactive query expansion. SIGIR’09 Proceedings of the 32th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, Boston, USA (2009) 17. Dumais, S., Cutrell, E., et al.: Stuff I’ve Seen: A system for personal information retrieval and re-use. Proceedings of SIGIR. (2006) 18. Morita, M., Shinoda, Y.: Information filtering based on user behavior analysis and best match text retrieval. ACM Conference on Research and Development in Information Retrieval (SIGIR) (1994) 19. Joachims, T., Granka, L., et al.: Accurately interpreting clickthrough data as implicit fe. SIGIR (2005) 20. Voorhees, E.: Evaluation by highly relevant documents. SIGIR’2001 Proceedings of the 24th ACM-SIGIR Conference on Research and Development in Information Retrieval, New York (2001) 21. Buckley, C., Voorhees, E.M.: Retrieval System Evaluation. TREC: Experiment and Evaluation in Information Retrieval. MIT Press (2005) 22. Jarvelin, K., Kekalainen, J.: IR evaluation methods for retrieving highly relevant documents. Proceedings of the 23rd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (2000)
Chapter 27
Example Verification of a Safety Evaluation Model for Road Passenger Transportation Fujia Liu, Chaozhou Chen, and Guoliang Dong
Abstract The development of network technology makes it possible for transportation companies to quickly collect and integrate information during transportation. The information extracted from different information management systems can be used for risk assessment and safety warning of transportation vehicles. Using the safety evaluation model can evaluate the risks in the transportation process in advance and prevent accidents. The quality of the evaluation model will affect the objectivity of the conclusion. This article introduces the evaluation indicators for safety evaluation. Aiming at the problem that the evaluation conclusions in the traditional evaluation model do not match the actual situation, an improved model method is proposed. The two actual cases were analyzed and compared with the evaluation model before and after the improvement. It was found that the improved evaluation model solved the problems of the traditional evaluation model, and the evaluation conclusions were consistent with the actual situation.
27.1 Introduction Once an accident occurs on a road passenger transport vehicle, it will cause great harm to the occupants and society. Evaluation of the safety status of the transportation process before the departure of the vehicle can predict risks in advance. Network management technology is widely used in the transportation process, including driver information management, vehicle technical status and maintenance information management, wireless and network dynamic monitoring, and many other aspects. The information in the information system and wireless transmission is input into the risk evaluation model, and the risk of the transportation process can be evaluated. The quality of the evaluation model determines the quality of the safety evaluation. Inappropriate evaluation models will lead to large deviations in evaluation results, which cannot meet the safety management requirements of transportation companies. Based on the study of the classic evaluation model, an improved transportation F. Liu (B) · C. Chen · G. Dong Research Institute of Highway Ministry of Transport, Beijing 100088, China e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_27
223
224
F. Liu et al.
process safety evaluation model is proposed. Using the evaluation model before and after the improvement to analyze and compare two actual cases, it is found that the improved evaluation model solves the problems in the traditional evaluation model, and the evaluation conclusion is consistent with the actual situation.
27.2 Evaluation Index The safety of transportation involves many factors such as drivers, vehicles, roads and environment, management, and so on. Accident statistics show that 85% of accidents are caused by the fault of the driver. Speeding and driving in violation of safe operating regulations are the main manifestations of driver faults. Drivers’ weak safety awareness leads to serious violations which are one of the main reasons for road traffic accidents [1–3]. The driver’s safety awareness determines the standardization of his driving behavior and safety awareness affects risk perception and driving skills [4]. The relationship between driving experience and safety is not linear [5–7]. Drivers who have been driving for less than 5 years have the greatest safety risk. Drivers younger than 25 years old have the greatest risk of driving safety. Drivers under 30 and over 57 are at higher risk. Drivers in the 35–40 and 49–54 age groups have lower safety risks. For those who have been driving for 5–10 years, the safety risks gradually increase. If the driving experience is more than 10 years, the risk gradually decreases. The number of accidents caused by mechanical failure is relatively small [7]. The key vehicle components that affect vehicle safety mainly include: brakes, steering, tires, seat belts, etc. Road alignments such as long slopes, steep slopes, and roads near rivers will have a greater impact on driving safety. In particular, conditions such as slippery roads, water accumulation, and icing caused by rain and snow weather are more likely to cause accidents under the above road conditions. The safety management of vehicles and drivers by transportation companies can improve the safety awareness of drivers and ensure that the technical conditions of the vehicles are in good condition. The regular safety education of drivers by transportation companies will significantly reduce accidents [8].
27.3 Evaluation Model 27.3.1 Evaluation Method The fuzzy comprehensive evaluation method used for the safety evaluation of the transportation process integrates the advantages of the analytic hierarchy process and has the advantages of simple calculation and reasonable evaluation results [9]. In the evaluation, it is necessary to calculate the membership degree subset for each evaluation index. For qualitative and quantitative indicators, different methods should be
27 Example Verification of a Safety Evaluation Model …
225
Fig. 27.1 Membership function based on the Cauchy distribution
used [10]. Indicators such as driving style, maintenance conditions, weather conditions, and road conditions are all qualitative indicators, and an expert method can be used to construct a subset of membership degrees. For quantifiable indicators, the evaluation score of the indicator can be obtained according to the scoring rules, and then, the membership function is used to convert the evaluation score into a subset of membership. Driving age, age, vehicle age, vehicle mileage, etc., are quantitative indicators. The assignment method is used to determine the membership function in the safety evaluation of the transportation process. Based on the Cauchy distribution in the safety evaluation of transportation process, the membership function curve of each evaluation level is determined, as shown in Fig. 27.1.
27.3.2 Traditional Evaluation Model The evaluation index is divided into three layers. The evaluation model requires that the values of the indicators in the sub-layer should be mutually exclusive, and the sum of the weight values of all the indicators in the sub-layer should be 1. The vectors A, R, and B are used to represent the weight value, evaluation matrix, and membership degree of each layer. The model of each layer index is as follows: The third layer:
226
F. Liu et al.
⎧ ⎪ ⎪ A11 = (a111 , a112 , a113 ) ⎪ ⎪ ⎪ B11 = A11 .R11 ⎪ ⎪ ⎨ A12 = (a121 , a122 , a123 , a124 , a125 , a126 ) ⎪ B12 = A12 .R12 ⎪ ⎪ ⎪ ⎪ A13 = (a131 , a132 , a113 , a134 , a135 , a136 , a137 ) ⎪ ⎪ ⎩ B13 = A13 .R13 ⎧ ⎪ A21 = (a211 , a212 ) ⎪ ⎪ ⎪ ⎪ B21 = A21 .R21 ⎪ ⎪ ⎨ A22 = (a221 , a222 , a223 , a224 , a225 ) ⎪ B 22 = A22 .R22 ⎪ ⎪ ⎪ ⎪ A 23 = (a231 , a232 , a233 ) ⎪ ⎪ ⎩ B23 = A23 .R23
A31 = (a311 , a312 , a313 , a314 , a315 , a316 , a317 , a318 ) B31 = A31 .R31
(27.1)
(27.2)
(27.3)
The second layer: ⎧ ⎨ A1 = (a11 , a12 , a13 ) R = (B11 , B12 , B13 )T ⎩ 1 B1 = A1 .R1 ⎧ ⎨ A2 = (a21 , a22 , a23 ) R = (B21 , B22 , B23 )T ⎩ 2 B2 = A2 .R2 ⎧ ⎨ A3 = (a31 , a32 , a33 , a34 ) R = (B31 , R32 , R33 , R34 )T ⎩ 3 B3 = A3 .R3
⎧ ⎪ ⎨ A4 = (a41 , a42 , a43 , a44 , a45 ) T R4 = R41 , R42, , R43 , R44 , R45 ⎪ ⎩ B = A .R 4 4 4
(27.4)
(27.5)
(27.6)
(27.7)
The first layer: ⎧ ⎨ A = (a1 , a2 , a3 , a4 ) R = (B1 , B2 , B3 , B4 )T ⎩ B = A.R Evaluation result k:
(27.8)
27 Example Verification of a Safety Evaluation Model …
⎡ ⎢ ⎢ k = B.C = (b1 , b2 , . . . , bn ) · ⎢ ⎣
c1 c2 .. .
227
⎤ m ⎥ ⎥ = bj · cj ⎥ ⎦
(27.9)
i=1
cn According to the section to which the evaluation score belongs, the risk level of this transportation process can be determined.
27.3.3 Improved Evaluation Model When the traditional evaluation model is used to evaluate the safety of the actual transportation process, there are often problems that the evaluation results are inconsistent with the actual situation. After analyzing the calculation process of the evaluation model, the main reasons are found. (1)
(2)
The indicators in each sub-layer of the conventional evaluation model are generally mutually exclusive. However, a single option cannot objectively express the actual situation. Therefore, for some special sub-layers, it should be allowed to check multiple items under the sub-layer. The traditional evaluation model stipulates that the sum of the weights of all evaluation indicators under the sub-layer should be 1. In the process of merging and superimposing indicator items, the degree of influence of highweight indicators will be weakened. Especially when multiple main indicators are concentrated under the same sub-layer, the degree of weakening of the importance of these indicators is even more serious. In the improved model, if the indicators under the sub-layer are not mutually exclusive, the sum of the weights of all indicators under the sub-layer no longer has to be 1.
27.4 Comparison Calculation of Two Examples In order to verify the evaluation effect of the improved evaluation model, two actual cases were evaluated to compare the risk of the evaluation model before and after the improvement. One of the cases was a traffic accident, and the other was a safe driving record of a company.
27.4.1 Example 1: A Traffic Accident In an accident in 2018, a bus collided with a small bus. Many passengers on the bus were thrown out of the bus, resulting in 8 deaths. The bus was 5.5 years old, the
228
F. Liu et al.
tires were severely worn, and some of the seat belts were damaged. The bus driver is 33 years old, 14 years of driving experience, and 13 illegal records. The accident section was a straight section, and the road surface was smooth. It was light rain when the accident happened. The road is wet. No water, no fog. The passenger transport company did not perform well in monitoring the vehicles. The driver’s “physical fitness” is calculated as “average,” with an evaluation score of 75 points. The driving experience is 14 years, and the evaluation score is 72 points. The age is 33 years old, and the evaluation score is 74 points. The vehicle age is 5.5 years, and the score is 77.5 points.
27.4.1.1
Traditional Method for Evaluation
The following are the calculation steps of the sub-layer of “driver.” ⎡
R11
⎤ 0.0190.2450.4900.2450.000 = ⎣ 0.0290.3320.4540.1840.000 ⎦ 0.0220.2710.4850.2220.000
B11 = A11 .R11 = (0.0240.2860.4760.2160.000) B12 = A12 .R12 = (0.0520.3230.4110.1940.000) B13 = (0.0000.0500.0910.2020.657) B21 = A21 .R21 = (0.0270.2900.4150.2230.045) B22 = (0.0050.0710.1680.3830.373) B23 = A23 .R23 = (0.0130.1750.3560.2940.162) B31 = (0.0190.2450.4900.2450.000) The evaluation matrix of the driver index is: ⎡ ⎤ 0.0240.2860.4760.2160.000 R1 = ⎣ 0.0520.3230.4110.1940.000 ⎦ 0.0000.0500.0910.2020.657 The membership value of the driver index is: B1 = A1 .R1 = (0.0220.1880.2740.2020.308) Using the same method, the membership degrees of the other indexes are: B2 = A2 .R2 = (0.0180.1660.2960.3080.210) B3 = A3 .R3 = (0.0130.1650.3490.3200.155) B4 = A4 .R4 = (0.1600.1400.1360.2750.289)
27 Example Verification of a Safety Evaluation Model …
229
The evaluation matrix of the first level indicators is: ⎤ ⎡ ⎤ 0.0220.1830.2680.2060.316 B1 ⎢ B2 ⎥ ⎢ 0.0190.1690.3000.3060.204 ⎥ ⎥ ⎢ ⎥ R=⎢ ⎣ B3 ⎦ = ⎣ 0.0130.1650.3490.3200.155 ⎦ B4 0.1600.1400.1360.2750.289 ⎡
The membership degree of the first level index is: B = A.R = (0.0520.1690.2560.2590.261) The risk evaluation score of this transportation process calculated by the traditional evaluation method is: k = B.C =
mn
b j · c j = 80.1
i=1
Therefore, the traditional evaluation method is used to evaluate the safety risk of the transportation process before the accident with a score of 80.1, which lies in the critical zone of “higher” risk and “normal” risk. The evaluation conclusion contradicts the actual situation.
27.4.1.2
Improved Method for Evaluation
Using an improved evaluation method for calculation, the safety risk evaluation score of the transportation process before the accident is 95.3 points, and the risk level is “very high.” The evaluation conclusion is consistent with the actual situation.
27.4.2 Example 2: A Safe Transportation Process In December 2019, a bus safely completed a long-distance transportation task. The driver is 48 years old and has been driving for 24 years. He has violated the law for speeding below 10% once in the past year. The vehicle is 7 years old. The key components are normal and maintained in good condition. The road passed by is normal and straight. The transportation has a good management of vehicles and drivers. The driver is 48 years old, and the evaluation score is 60 points. The driving experience is 24 years, and the evaluation score is 60 points. The driver’s “physical fitness” is “good,” and the evaluation score is 65 points. The driver is “familiar” with the driving route and scores 65 points. The driving style is “cautious” and scores 60 points. The driving technique is “very good,” with a score of 55 points. The driver’s
230
F. Liu et al.
condition before departure was “very good,” with a score of 55 points. The driver violated the law once and scored 75 points. The vehicle is 7 years old and scores 85 points. The risk evaluation score of this transportation process calculated by the traditional evaluation method is 68.3 points, which lies in the critical zone of “lower” risk and “normal” risk. The evaluation conclusion is consistent with the actual situation. The risk evaluation score of this transportation process calculated by the improved evaluation method is 67.0 points, and the risk level is “low.” The evaluation conclusion is consistent with the actual.
27.4.3 Comparative Analysis For the above two cases, the evaluation models before and after the improvement were used to calculate the individual evaluation values of driver, vehicle, road and environment, and management factors. The evaluation results are shown in Tables 27.1 and 27.2. It can be seen from Table 27.2 that the overall evaluation results of accident cases using the model before and after the improvement are quite different (19.3%). The calculation results of the two models of the “vehicle” indicator and the “management factor” indicator differ greatly, which are 49.4% and 15.4%, respectively. The “vehicle” indicator has cancelled the requirement that the sum of the weight indicators within the layer is 1, and the single evaluation result is 119.4, breaking the limit of 100. Compared with the result calculated using the traditional evaluation Table 27.1 Evaluation results of an accident case Items
Improved
Overall evaluation Individual evaluation
Driver Vehicle Road and environment
Difference (%)
95.3
79.9
84.9
80.4
5.5
79.9
49.4
119.4 82.3
Management
Traditional
91.2
19.3
79.5
3.6
79.0
15.4
Table 27.2 Evaluation results of a safe transportation Items
Improved
Difference (%)
67.0
68.2
−1.7
Driver
67.2
69.0
−2.5
Vehicle
69.2
67.9
2.00
Road and environment
78.5
77.0
2.0
Management
59.1
63.3
−6.7
Overall evaluation Individual evaluation
Traditional
27 Example Verification of a Safety Evaluation Model …
231
method (79.9 points), the calculated result using the improved model (95.3 points) objectively reflects the degree of risk in this transportation process. It can be seen from Table 27.2 that the overall evaluation results of accident cases using the model before and after the improvement are not much different (1.7%). The calculation results of the “management factor” differ by 6.7%. The calculation results of other indicators differ by less than 2.5%. The evaluation model before and after the improvement has little difference for the safe transportation process.
27.5 Conclusion The extensive application of information technology in the transportation process has improved the efficiency of data collection and can provide a large amount of useful information for the risk assessment of the transportation process. The analysis of these multi-source data can realize the risk assessment of the transportation process. This paper introduces the evaluation indicators for the safety evaluation of road passenger transportation. The reasons for the inconsistency between the traditional evaluation model and the actual situation are analyzed, and an improved evaluation model program is proposed. The evaluation model before and after the improvement is applied to evaluate two actual cases. After comparing the evaluation results, it is found that the traditional evaluation model can basically reflect the status of the safe transportation process, but there is a big deviation in the evaluation of the dangerous transportation. The improved evaluation model for safe transportation process and dangerous transportation process evaluation results is consistent with the actual. The comparative results of case analysis show that the improved evaluation model is effective. Acknowledgements This paper was funded by the 2020 Basic Scientific Research Business Expense Project of Research Institute of Highway Ministry of Transport, “Research on the Technical Management Methods of Road Transport Vehicles Based on Big Data” (2020-9034).
References 1. Wu, B.J., Sun, Q., et al.: Analysis of key influencing factors of automobile transportation safety management. Logistics Eng. Manag. 040(001), 158–160 (2018) 2. Zhang, Q.: Research on driver safety awareness assessment, Ph.D. Thesis, Chang’an University (2010) 3. Gao, L.N.: Research on evaluation method of driver’s safety awareness simulation driving, Ph.D. Thesis, Kunming University of Science and Technology (2015) 4. Tao, J.: In-depth analysis of drivers of commercial vehicle traffic accidents. Shanxi Transp. Sci. Technol. 253(04), 155–158 (2018) 5. Yuan, L., Yuan, H.W., et al.: Analysis of the influence of driver’s age on traffic risk perception ability. Soc. Psychol. Sci. 6, 91–94 (2013)
232
F. Liu et al.
6. Shi, N.N.: Analysis of the frequency distribution of driver accidents and its influencing factors. Beijing Jiaotong University, Ph.D.thesis (2018) 7. Zhu, S.S.: Research on the traffic safety risk of operating buses based on the accident tree analysis, Ph.D. Thesis, Chang’an University (2016) 8. Wang, X.S., Luo, Y., et al.: Traffic safety education for operating drivers based on risky driving behavior analysis. Automob. Safety 1, 62–66 (2016) 9. Wu, Z.J., Fan, W.J.: Research on risk assessment system of road transportation of dangerous goods. Highw. Transp. Sci. Technol. 2(12), 6–11 (2015) 10. Xiao, W.: Research on transportation safety evaluation of automobile units, Ph.D. Thesis, Nanjing University of Science and Technology (2009)
Chapter 28
Development of Broadband Sensors Based on Fiber Optic Gyroscope Bo Yang, Yong Li, Yafei Zhao, Peiyong Zhang, Li Sun, Xin Hui, and Xiaona Zhou
Abstract Due to the observations of out small disturbance in spacecraft, a promising new field of research in inertial attitude sensor has been developed. However, exploration of new spacecraft under high-frequency oscillation has been hampered by the lack of a portable, reliable and highly sensitive broadband sensor. In this paper, the laboratory tests of a broadband inertial attitude sensor named micro-vibration measurement sensor (MMS) in spacecraft based on interferometric fiber optic gyroscope (IFOG) is firstly reported. The noise level by operating range diagrams and Allan deviation is estimated and tests of the sensors, which are demonstrated the reliability of MMS in filed installations. This new technology presents a wide range of field applications in spacecraft, earthquake engineering and seismology.
28.1 Introduction For last few years, inertial attitude sensor (INS) has been looking for a reliable and highly broadband sensitive sensor for application in spacecraft which need highfrequency attitude information in the field of spacecraft attitude control with high stability [1–3]. In theory, interferometric fiber optic gyroscope (IFOG) has a high bandwidth due to its principle and will play an important role in obtaining attitude information. Therefore, well-known institutions at home and abroad have invested a lot of energy to develop a new production in this field [4–6]. For example, three SRS-2000 fiber optic gyroscopes from Optolink is the cutting edge in inertial rotational velocity √ measurements with the Bias drift of 0.003°/h and the random walk of 0.0002°/ h, which is shown in Fig. 28.1 [7]. This device has a bandwidth of 50 Hz, which can also be triggered at up to 2 kHz. This kind of device can be used to estimate the attitude in the inertial reference frame. And it has been first launched in Fall 2016 at National Aeronautics and Space Administration (NASA) Goddard Space Flight Center for the purpose of studying galactic clustered star formation by providing spatially resolved spectroscopy of nearby star clusters. B. Yang (B) · Y. Li · Y. Zhao · P. Zhang · L. Sun · X. Hui · X. Zhou Beijing Institute of Control Engineering, Beijing 100094, China e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_28
233
234
B. Yang et al.
Fig. 28.1 SRS-2000 and Astrix 200 fiber optic gyroscope
(a)SRS-2000 fiber optic gyroscope
(b)Astrix 200 fiber optic gyroscope
The Astrix200 is the highest precision fiber optic gyroscope for a mission led by the American, French, Canadian, and United Kingdom Space Agencies (UKSA), NASA, Centre National d’Etudes Spatiales (CNES), which launched in 2020 to study Earth’s surface water, which shown in Fig. 28.1. This device has been flying on Pleades Satellites and the flight experience of 13 satellites effectively verified the reliability and performance of the products. Nowadays, with the increasing demand of high resolution and high-reliability satellite platform for high-speed attitude measurement, it is very important to achieve stable control of satellite, while obtaining as much attitude information as possible. In this paper, a broadband micro-vibration measurement sensor (MMS) which can be used in satellite control is introduced.
28.2 Instrument and Method The technology reported in this work is based on a closed-loop interferometric fiber optic gyroscope, which means that its operating point is kept at the point of the highest sensitivity by an internal feedback loop. And the diagram of the proposed MMS shown in Fig. 28.2, which consists of optical circuit and circuit. In our scheme, after being amplified and filtered, the weak signal received by the detector is converted into digital signal by analog-to-digital converter (ADC), and then output to digitalto-analog converter (DAC) after being processed by field programmable gate array
28 Development of Broadband Sensors Based on Fiber …
235 MIOC
Coupler ASE
Coil
Optical circuit
PIN
Constant Current Driver
Amplifier Temperature Control
Driving Circuit FPGA
Parallel ADC
Circuit
Parallel DAC
Fig. 28.2 Schematic diagram of the single axis IFOG
(FPGA) and loaded on multi-functional integrated optical chip (MIOC) after being filtered by differential filter, which is used to realize closed-loop control. Therefore, in order to achieve the purpose of high precision and high bandwidth, the optical path and circuit are optimized in detail.
28.3 Configuration of Space Grade Single Axis IFOG 28.3.1 ASE Light Source In order to improve the signal-to-noise ratio of the light signal arriving at the detector, an amplified spontaneous emission (ASE) light source with a power of 5 mW is carried out, of which the scheme is shown in Fig. 28.4. In this scheme, the ASE light source including pump laser (PUMP), wavelength division multiplexer (WDM), erbium-doped fiber and fiber reflector (FR). The length and erbium ion concentration are carefully designed and optimized to meet the needs of high power and high stability. The experimental results show that all the parameters meet the application requirements of fiber optic gyroscope. The spectral type and specific test indicators are shown in Fig. 28.3 and Table 28.1, respectively. -10
-20
WDM Erbium-doped fiber
-30
Output
Power(dBm)
DUMP
-40
-50
-60
FR
-70 1520
1530
1540
1550
1560
1570
Wavelength(nm)
Fig. 28.3 The scheme of ASE light source (Left: Scheme; Right: Gauss spectroscopy)
1580
1590
1600
236
B. Yang et al.
Fig. 28.4 Schematic diagram of fiber optic coil and test curve
Table 28.1: Parameters of ASE light source (@100 mA)
Parameter
results
Power, mW
≥5
Average Wavelength, nm
1530
3 dB Spectral Width, nm
≥10
Ripple, dB
≤0.2
Spectral type
Gaussian
28.3.2 Fiber Optic Coil In recent years, different winding methods of fiber optic coil show different characteristics for the environmental adaptability of IFOG. The fiber optic coil with good
28 Development of Broadband Sensors Based on Fiber …
237
symmetry is of great significance in reducing the influence of temperature gradient for application of high precision. At present, the octupole symmetrical winding method is usually used in the optic fiber coil encircling system, which has a certain impact on the performance of high-precision IFOG due to the residual temperature error. The octupole-winding method which shown in has more obvious advantages and can further reduce the influence of temperature residual. Figure 28.4 shows the temperature characteristic curve of fiber optic coil for octupole-winding. It is found that when the temperature range is –10 to 50 °C, the fiber optic coil with octupole-winding has obvious advantages, and the bias stability over –10 to 50 °C is around 0.021°/h.
28.3.3 MMS Performance In order to enhance the performance of MMS, fundamental limitation is reduced as much as possible during our process. It can be summarized as follows: Octupole-winding technology is used to further improve the environmental adaptability of the fiber ring with 3.8 km. The whole machine structure thermal design is convenient to improve the product symmetry and reduce the influence of temperature gradient. The software algorithm based on FPGA is optimized, and the bandwidth is up to 2 kHz. Software filtering and hardware filtering are combined to reduce the noise of the product. After optimization, single axis IFOG with different precision is designed to fulfill the requirement of as more as more missions by our team. For example, FOG-1–1 series single IFOG will be used in China’s Mars exploration due to its miniaturization. FOG-2-1 series single axis IFOG has important application value in the field of commercial satellite and communication satellite, and has been widely verified in “Superview-1” satellite. In the early stage of design, considering the application requirements of remote sensing satellite, FOG-3-1 series single axis IFOG has carried out the optimization of high bandwidth and high precision, and the communication frequency can reach up to 2 kHz. The Allan variance curves of the three single axes IFOG are shown in Fig. 28.5. Aiming at the requirement of different spacecraft, different series of MMS products are developed after optimization, and the main indicators are shown in Table 28.2. All the single axis gyros can be used in the aerospace field and can be configurable for different configurations as 3S, 3S + 1 and 6S structure.
238
B. Yang et al.
Fig. 28.5 Single Axis fiber optic gyroscope (Allan variance)
Table 28.2: Parameters of single axis series Parameter
Single axis FOG-3-1 Single axis FOG-2-1 Single axis FOG-1-1
Range of measured angular ±20 rate, °/s
±110
±200
Bias drift, °/h
0.001
0.01
0.1
Scale factor repeatability, ppm
10
50
100
Bandwidth, kHz
Up to 2
Up to 2
Up to 2
0.0002
0.001
0.01
Weight, kg
1.2
0.42
0.28
Dimensions, mm
120 × 56
98 × 98 × 52
70 × 70 × 61.5
√ Random walk, °/ h
28.4 Conclusion In conclusion, development of broadband sensors based on fiber optic gyroscope is carried out and further optimization will be carried out to fulfill the requirement of specific customer demands. Compared with other countries’ products, our products have maintained a high-cost performance, and will play an increasingly important role in China’s aerospace missions in future. Acknowledgements This work was supported by a project numbered 41417060102.
28 Development of Broadband Sensors Based on Fiber …
239
References 1. Ge, S., Yang, R., Guo, C.: Constant small tension control for fiber optic coil variable-velocity winding. IEEE Access 7, 172012–172020 (2019) 2. Chen, S., Lin, T., Shih, Y., Hsu T.: Dynamic characteristic modeling and simulation for a fiberoptic gyroscope integration system. In: 2019 International Conference on Engineering, Science, and Industrial Applications (ICESI), pp. 1–6, Tokyo, Japan (2019) 3. Tehrani, M.M., Soltanieh, S.: Sources of bias instabilities (1/f Noise) in optical gyros. IEEE Sensors J. 22(19), 10554–10559 (2019) 4. Wang, H., Lu, Y., Lu, T., Liu, R., Liaw, S.: Superfluorescent fiber source with ultra-low thermal coefficiency operating in the conventional band. In: 2018 Conference on Lasers and ElectroOptics Pacific Rim (CLEO-PR), pp. 1–2 (2018) 5. Yao, J., Li, K., Li, B., Wang, C., Chen, K., She, X.: Analysis of spectrum properties of integrated optical chips applied on IFOG. In: 2018 Progress in Electromagnetics Research Symposium (PIERS-Toyama), pp. 827–832 (2018) 6. Gao H., Gao W., Wang G., Sun J.: Research on optimum modulation phase of interferometric fiber optic gyroscope in the space radiation. 2018 DGON Inertial Sensors and Systems (ISS), pp. 1–11 (2018) 7. Maxime, J.R., Rinehart, S.A., Dhabal, A., et al.: The Balloon Experimental Twin Telescope for Infrared Interferometry (BETTII): towards the first flight, Publications of the Astronomical Society of the Pacific, vol. 941 (126) (2016)
Chapter 29
Short-Term Wind Speed Prediction Based on Phase Space Reconstruction and Elman Network-Grey Wolf Optimization Shengqing Li, Haotian Feng, and Xi Deng Abstract Aiming at the problem of low accuracy of short-term wind speed prediction caused by sudden changes in wind speed. A combined short-term wind speed prediction algorithm based on the phase space reconstruction method and the gray wolf algorithm to optimize the Elman network is proposed. Firstly, the algorithm uses the phase space reconstruction method to reconstruct the original sequence data, and it constructs the time series prediction model. Then, the Elman network is used to train the model and the gray wolf algorithm is used to optimize the network parameters. Finally, evaluation index effected by the prediction is proposed to evaluate the prediction effect. Through simulation and testing, this method can better improve the prediction accuracy.
29.1 Introduction 29.1.1 Background With the increasing depletion of fossil energy and the severe challenges posed by environmental problems, wind power as a clean and sustainable energy source, is becoming the main future source of energy. However, due to the change of wind power, its large-scale interconnection easily leads to the stability decline of grid voltage, frequency, power angle, and so on, which would bring potential safety hazards to production and daily life [1–3]. By predicting the future wind power ahead of time, it can control the predicted power ahead of time and making the generation plan. Power predication of short-term wind can effectively prevent power grid stability problems that are caused by power fluctuation of wind. Wind power prediction research now is basically mature. There are two main predication methods: physical method and statistical method [4]. Firstly, the physical method establishes the meteorological model of the wind power plant by investigating the physical factors, such as, topography, climate background and so on. Then S. Li · H. Feng (B) · X. Deng Hunan University of Technology, Zhuzhou 412007, Hunan, China © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_29
241
242
S. Li et al.
the meteorological information is predicted by the meteorological model. Finally, the output power of the fan is obtained by taking the meteorological information as an independent variable and the output curve of the fan as a function [5]. Methods in statistics is based on statistical theory. Firstly, historical operation data of wind power plant are collected including historical wind speed and historical power. Then the mapping relationship between them is analyzed. Finally, future power is predicted by learning this mapping relationship. There are several methods employed to analyze this mapping relationship such as time series method [6], Kalman method [7], machine learning [8, 9], support vector machine [10–17], grey prediction [18, 19]. These methods have their advantages and disadvantages, and the prediction accuracy is also different [20, 21]. This paper solves the problem of improving the accuracy of the wind power forecasting. Combined forecasting algorithm based on Elman network-grey wolf optimization is proposed. And its effectiveness is verified by the modeling and analysis of MATLAB 2015.
29.1.2 Phase Space Reconstruction Nowadays, wind power plant can only collect meteorological information for example wind speed, wind direction, air pressure and temperature. However, for non-linear systems, the time series of one or several components can not meet the input requirements of the system. Besides, the degree of information correlation in time dimension is not considered, which often results in the loss of input information. On account of this, the input of the system can be reconstructed to increase the input information saturation. Embedding dimension h and time delay p are two important parameters for phase space reconstruction. Two important indicators of phase space reconstruction are the standard embedding dimension h and the time delay p. Based on the appropriately determined values of h and p, if the time series of wind speed is {x(t)|t = 1, 2, …, N}, the reconstructed wind speed sequence would be Y (t) = [x(t), x(t + p), x(t + 2 p), ..., x(t + (h − 1) p)], t = 1, 2, 3, ..., N − (m − 1) p
(29.1)
Among them, G-P (Grassberger-Procaccia) algorithm is used to determine the value of h. And p is usually determined by autocorrelation function method, average mutual information method and reconstruction expansion method. The pseudo-phase space is composed of Y and is recorded as Rm . The evolution rule of Rm midpoint Y (t) → Y (t + p) is the same as that of the original system, and many features in that system can be reflected in the reconstruction space. Formula (29.1) is taken as input and Y (t + p) as output to establish prediction model. The mapping relationship between input and output is determined to realize wind speed prediction.
29 Short-Term Wind Speed Prediction …
243
29.2 Elman Network A typical recurrent neural network calls elman neural network (RNN), as shown in Fig. 29.1. Its internal feedback and feedforward structure’ equips it with stronger learning ability than ordinary neural network. It is very suitable for building a nonlinear model of wind speed prediction. In the learning algorithm of Elman network S(k) and y(k) are as follows: S(k) = f (W1 X + W3 S(K − 1) + B1 ) y(k) =
W2 S(k) + B2
(29.2) (29.3)
In formula (29.2)–(29.3), k represents time; f denotes the transfer function of the hidden layer and generally takes S-type function; S(k − 1) is the feedback term of the hidden layer. The offset values of the input layer B1 and the hidden layer B2 . The weights of the input layer W 1 , the hidden layer W 2 , and the output layer W 3 . The input values X and the output values y. The global approximation error is: E=
T 1 [y(k) − y 0 (k)]2 2 K =1
(29.4)
In formula (29.4): y0 (k) represents the target output at k time. The modification of weights W 1 and W 2 is similar to that of feedforward neural networks. The main difference lies in the weight W 3 . In reference [10], a chain rule is proposed to obtain the weights, e.g. (29.5):
Fig. 29.1 Elman neural networks structure
244
S. Li et al.
W3 =
T
[(y 0 (k) − y(k))W2 ] f (S(k))S(k − 1)
(29.5)
k=1
29.3 Grey Wolf Algorithm 29.3.1 Rationale of Grey Wolf Algorithm The main wolf is marked as alpha (α), and the remaining grey wolves are recorded as beta, delta and omega (β, δ and ω) in the algorithm. The alpha wolves are always in the best fitness position and the closest to the optimal value, while the beta and delta as the candidates for alpha wolves have their second best fitness, and the rest are omega wolves. In the whole algorithm process, alpha wolves are responsible for leading wolves to search, track and approach prey, beta and delta wolves attacking prey, and omega wolves hunting prey.
29.3.2 Description of Grey Wolf Algorithm The algorithm can divide the whole hunting process into three steps: encircling, chasing and attacking, and ultimately capture the prey (perform the optimization). Encircling: The distance between prey and wolf during this period.
D = C ∗ X p (t) − X (t) X (t + 1) = X p (t) − A ∗ D
(29.6)
The distance between prey and gray wolf D, and the number of iterations t, respectively, given by formula (29.6). The position vector of prey X p (t) (i.e. the location of the optimum solution) and the position vector of grey wolf X(t), obviously, given by formula (29.6) after the iteration (i.e. the location of the latent solution). The coefficients A and C, and their formulas are as follows: A = 2m ∗ r1 − m (29.7) C = 2r2 When the number of iterations increases, and m linearly decreases from 2 to 0, r 1 and r 2 are random numbers between [0, 1]. Chasing: The beta and delta wolves are led by the alpha wolves to hunt their prey. The position of the prey (optimal solution) is re-determined. The renewal equation is as follows:
29 Short-Term Wind Speed Prediction …
245
⎧ ⎪ 1 X α (t) − X (t)| ⎨ Dα = |C Dβ = C2 X β (t) − X (t) ⎪ ⎩ Dδ = |C3 X λ (t) − X (t)| ⎧ ⎪ ⎨ X 1 = X α (t) − A1 Dα X 2 = X β (t) − A1 Dβ ⎪ ⎩ X 3 = X λ (t) − A1 Dλ X p (t + 1) =
X1 + X2 + X3 3
(29.8)
(29.9)
(29.10)
Among them, Dα , Dβ , Dλ represent the distance between the wolves α, β, λ and the wolves ω (other individuals). Attacking: It is realized by decreasing the value of a in Eq. (29.10). When A are in [−1, 1], it indicates that wolves would converge towards the prey, while when 1 < |A| ≤ 2, wolves would diverge from the prey, which would deprive the grey wolf algorithm of its optimal position, thus falling into a local optimal process.
29.3.3 Grey Wolf Optimization Algorithm See Fig. 29.2.
29.4 Wind Speed Combination Predication and Analysis 29.4.1 Prediction Process The traditional wind speed prediction model takes the single-dimensional wind speed data as the input, which causes a great loss of implicit information. However, the wind speed in the adjacent time period would affect the trend of its future wind speed change. By means of autocorrelation method and G-P method, this paper takes the maximum value of future wind speed impact factor as the objective function to investigate and solve the influence of historical wind speed on future wind speed in different phases. The reconstructed wind speed input can not only effectively release the hidden information in the wind speed sequence, but also maximize the learning ability of Elman network with appropriate dimension input. In order to effectively improve the learning ability and learning rate of the Elman network, this paper uses the gray wolf optimization algorithm to optimize the initial weights and thresholds of the network. In the model training stage, through batch training process, the weights and thresholds of n groups are obtained, and they are all optimal weights and thresholds; then the weights and thresholds of n groups
246 Fig. 29.2 Grey wolf optimization algorithm
Fig. 29.3 Flow chart of combined model for wind speed prediction
S. Li et al.
29 Short-Term Wind Speed Prediction …
247
are cross-validated, and the optimal group of weights and thresholds is selected for the test stage. Through cross-validation, the generalization ability of weights and thresholds for different batches of inputs can be well tested, and a group of weights and thresholds with the strongest generalization ability can be screened as well. By optimizing the input, weight and threshold, a combined model of short-term wind speed prediction based on Elman network-Grey wolf optimization is proposed in Fig. 29.2 (Fig. 29.3).
29.4.2 Error Evaluation Function The estimation of prediction error is evaluated by three indexes: MSE, MAE and maximum absolute error. (29.11)
In the formula, n is the data value and x is the wind speed vector.
29.4.3 Simulation Experiment and Result Analysis The original data is the actual wind speed of a wind farm from October 1, 2016 to December 31, 2017. With 15 min as a sampling point, nearly 30,000 wind speed data were collected. After preprocessing, 16,213 pieces of available data are obtained, as shown in Fig. 29.4. From the wind speed curve, they cover all the wind speed sections, which can be regarded as a set of ideal simulation data to reflect the actual change.
Fig. 29.4 Wind speed of wind field
248
S. Li et al.
Fig. 29.5 Comparison of predicted wind speeds in different experiments
Input and Output: The original data is cleaned and reconstructed. As the final input of Elman network, p is determined to be 15 by autocorrelation method and h is determined to be 6 by G-P algorithm. Therefore, the output and input data of each wind speed are column vectors of 1 * 6, i.e. input and output is as formula (29.12):
X = [x(t), x(t + 15), x(t + 30), x(t + 45), x(t + 60), x(t + 75)] Y = x(t + 90)
(29.12)
Training phase: Take the first 15,000 cleaned wind speeds as training data, and use the Grey wolf algorithm to optimize the network. Test phase: Take 1000 cleaned wind speed as test data. Through three groups of different experiments, the prediction effect of the combined algorithm is compared and analyzed. Experiment 1 Without pretreatment of wind speed, that is, without phase space reconstruction, the input of Elman network becomes a single wind speed of 1 * 1. The output remains unchanged, and the weights and thresholds of Elman network are not initialized and optimized either. Experiment 2 Reconstruct the phase space of wind speed, take p = 15, h = 6, and employ the unoptimized Elman network to predict the reconstructed wind speed data. Experiment 3 Reconstruct the phase space of wind speed, optimize Elman’s initial weights and thresholds by using grey wolf algorithm, and predict the wind speed. Figure 29.5 is a comparison of the predicted and actual values of wind speed obtained from three groups of different experiments. From Fig. 29.5, it can be seen that with the preprocessing of input, coupled with the initialization of Elman network parameters by Grey wolf algorithm in the early stage, the prediction curve gradually approximates the actual value. The prediction curve is smooth, and when there is a significant fluctuation in wind speed, its prediction can also maintain a high stability.
29 Short-Term Wind Speed Prediction … Table 29.1 Comparison of three types of wind power forecasting errors
249
Experiment
MSE
MAE
One
0.4064
0.2506
Two
0.3042
0.1823
Three
0.2239
0.1279
According to Table 29.1, the MSE and MAE in experiment 3 are 0.2239 and 0.1279, which are the minimum error values of previous experiments. It shows that the wind speed prediction model based on Elman network-grey wolf optimization can more accurately predict the wind speed process, and the effect is better than the first two experiments. On the basis of the above simulation results, by reconstructing the appropriate input and using the optimized Elman network, which can enhance the accuracy of wind speed prediction. The combined model of short-term wind speed prediction based on Elman network-grey wolf optimization can greatly improve the accuracy of prediction.
29.5 Conclusion Aimed at the problem of poor prediction accuracy caused by the randomness of wind speed, the appropriate input is reconstructed by phase space, and combined with the optimized Elman network to predict wind speed, which greatly improves the prediction accuracy. The conclusions are as follows: (1) (2)
(3)
Data reconstruction can effectively display and magnify the hidden information of the data, and increase the information input of the prediction model. By optimizing the initial weights and thresholds of Elman network through grey wolf algorithm, the prediction speed and convergence rate of the network can be effectively enhanced and the prediction accuracy of the network can be reached as well. The combined model of short-term wind speed predication based Elman network-grey wolf optimization has better prediction accuracy than the single model, and MSE, MAE of prediction results are reduced as well.
References 1. Cheng, S.Y., Song, S.F., Li, L.X.: Survey on smart grid technology. Power Syst. Technol. 8, 40–43 (2009) 2. Lange, B., Rohrig, K., Ernst, B.: Wind power prediction in Germany-recent advances and future challenges. In: European Wind Energy Conferences, D-34119 Kassel, Germany (2016)
250
S. Li et al.
3. Lange, M., Focken, U., Heinemann, D.: Previento-regional wind power prediction with risk control. In: World Wind Energy Conference, Berlin Germany (2012) 4. Qian, Z., Pei, Y., Cao, L.X.: Review of wind power forecasting method. High Voltage Eng. 4, 1047–1060 (2016) 5. Zhu, X.R., Liu, Y.P.: Wind power forecasting using time series model based on robust estimation. Proc. CSU-EpSA 3, 103–110 (2012) 6. Zhou, S.L., Miao, M.Q., Su, J.W.: Prediction of wind power based on principal component analysis and artificial neural network. Power Syst. Technol. 4, 19–23 (2014) 7. Pan, D.F., Liu, H., Li, Y.F.: Optimization algorithm of short-term multi-step wind speed forecast. Proc. CSEE 26, 87–91 (2008) 8. Cui, Q., Jie, S., Zhang, X.: The application of improved BP neural network for power load forecasting in the island microgrid system. In: International Conference on Electrical & Control Engineering, Yichang China (2011) 9. Zhang, L.Y., Meng, L.J., Wang, Z.Z.: Photovoltaic power station output power prediction based on double Bp neural network. Electr. Meas. Instrum. 11, 31–35 (2015) 10. Li, R., Ke, Y.Q., Zhang, X.: G: Wind power forecasting based on time series and SVM. Electr. Power 1, 64–68 (2012) 11. Ding, Z.Y., Yang, P., Yang, X.: Wind power prediction method based on sequential time clustering support vector machine. Autom. Electr. Power Syst. 14, 131–135 (2012) 12. Wang, K., Guan, S.Q., Wang, L.X.: A combined forecasting model for wind power prediction based on fuzzy information granulation and least squares support vector machine. Power Syst. Prot. Control 2, 26–32 (2015) 13. Li, X., Wang, X., Zhen, Y.H.: Short-term wind load forecasting based on improved LSSVM and error forecasting correction. Power Syst. Prot. Control 11, 64–69 (2015) 14. Chen, D.J., Gong, Q.W., Jin, C.Y.: Short-term wind power prediction based on support vector regression machine optimized by adaptive disturbance quantum-behaved particle swarm optimization. Power Syst. Technol. 4, 974–980 (2013) 15. Yang, Z.L., Liu, Y.: The prediction of short-term wind power by applying particle swarm optimization. Power Grid Technol. 5, 159–164 (2011) 16. Wang, Y.X., Chen, G.C.: Short-term wind power prediction based on IAFSA optimization SVM. Electr. Meas. Instrum. 3, 80–84 (2016) 17. Ye, L., Liu, P.: Combined model based on EMD-SVM for short-term wind power prediction. Proc. CSEE 31, 102–108 (2011) 18. Wang, L.J., Dong, L., Liao, X.: Z: Short-term power prediction of a wind farm based on wavelet analysis. Proc. CSEE 28, 30–33 (2009) 19. Meng, X.X., Tian, C.W., Dong, L.: A study of grey theory used in prediction of medium and long-term wind power generation. Power Syst. Prot. Control 21, 81–85 (2011) 20. Meng, Y.F., Hu, S.J., Deng, Y.: Analysis and evaluation method of wind power prediction-error. Power Constr. 7, 6–9 (2013) 21. Li, K., Gao, Y.: A combination method for wind power prediction based on cooperative game theory. Electricity (Z1), 2, 36–40 (2014)
Chapter 30
Modern-Driven Deep Learning for Belief Propagation LDPC Decoding Yiduo Tang, Lin Zhou, Shuying Zhang, and Yan Chen
Abstract In this paper, for the long code decoding problem, we analyze the performance of belief propagation (BP) decoder in neural network. The decoding of long codes has always been a concern of LDPC decoding. In recent years, the application of neural networks in the communication field has gradually become widespread. As a result, we are considering and combining the two. The decoding method proposed in this paper uses model-driven deep learning. The network we propose is a neural standardized BP LDPC decoding network. Model-driven deep learning absorbs the advantages of both model-driven and data-driven, which combines them adaptively. The network structure proposed in this paper takes advantage of model-driven to expand the iterative process of decoding between check nodes and variable nodes into the neural network. We can increase the number of iterations by increasing the CN layer and VN layer of the hidden layer. Furthermore, by changing the SNR to detect its relationship with system robustness, and, finally, determine the appropriate SNR range.
30.1 Introduction LDPC codes which are a well-known linear error correcting codes were first proposed by Gallager in 1962 [1]. However, due to the limitation of computing power and hardware development level at that time, the code did not attract attention at that time. Since the 1996 literature [2] rediscovered LDPC codes, this code has attracted the attention of scholars together with Turbo codes due to its excellent error correction performance approaching the Shannon limit and its high decoding throughput and has become a new research hotspot. Y. Tang · L. Zhou · S. Zhang · Y. Chen (B) Xiamen Key Laboratory of Mobile Multimedia Communications, National Huaqiao University, Xiamen 361021, Fujian, China e-mail: [email protected] L. Zhou State Key Laboratory of Integrated Services Networks, Xidian University, Xi’an 710071, Shaanxi, China © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_30
251
252
Y. Tang et al.
So far, belief propagation (BP) decoding algorithm is one of the main LDPC decoding algorithms. The computational complexity of the BP decoding algorithm is related to the code length of the LDPC code. It can be applied in a variety of environments and can achieve good performance. Researchers have been conducting in-depth research and improvement on LDPC encoding and decoding algorithms. Deep learning has powerful learning and extracting feature attributes, which is suitable for dealing with high complexity problems. In recent years, researchers have gradually begun to use deep learning to solve decoding problem in communication. It has been illustrated that assigning appropriate weights to the transmission information in Tanner graph can reduce iterations of the BP decoder with identical decoding performance [3]. In [4], author proposed a decoding method which used neural min-sum to reduce the complexity of decoding in hardware. In order to design a decoder for LDPC long code with lower complexity, we propose a new neural BP LDPC decoding network. Iterative decoding is performed by unfolding the BP of long code decoding into a feedforward propagation network.
30.2 Related Work 30.2.1 BP Decoding LDPC codes are a kind of (N, K) linear block codes with good forward error correction performance. Among them, the parity check matrix H is sparse, and the number of non-zero elements in the matrix is much smaller than the number of zero elements. For (N, k) LDPC codes, the code length is represented by N, which value is equal to the total number of columns in the check matrix. M represents the length of the check sequence, which value also shows the total number of rows of the check matrix. K represents the information sequence length. In addition to using the check matrix to represent an LDPC code, the graph model is another way to represent LDPC codes. The Tanner graph is a common method in the graph model. The Tanner graph is only composed of check nodes, variable nodes, and the connection between the two, and the connections are all undirected [5], in which each check node is summed one row of the check matrix corresponds to each other, and each variable node corresponds to a column of the check matrix. Therefore, the Tanner graph includes M check nodes and N variable nodes in total, and when the element h in the ith row and the j column of the check matrix is h = 1, there is a line between the ith check node and the jth variable node in the Tanner graph. In the BP decoding process, the messages in the channel are transmitted iteratively using the variable nodes and check nodes of the Tanner graph. We use ri define channel output. Let L(r ji ) denotes the messages from CN to VN, L(qi j ) denotes messages from VN to CN, and L(Pi ) denotes the log-likelihood ratio (LLR) of the
30 Modern-Driven Deep Learning for Belief Propagation LDPC Decoding
253
receiving channel corresponding to bit i. The information delivery process is as follows: Step 1
The channel output signal ri send to the decoder and then decoder calculates the corresponding LLR, the nth coded bit wi as follow: L(qi j ) = L(Pi ) = log
Step 2
Pr(wi = 0|ri ) 2ri = 2 Pr(wi = 1|ri ) σ
where σ 2 is the channel noise power following the Gaussian distribution. Message updated in VN: ⎛ L(r ji ) = 2 tanh−1 ⎝
⎞ 1 tanh L qi j ⎠ 2 \i
i ∈R j
Step 3
(30.1)
(30.2)
Message updated in CN:
L (l) (qi j ) = L(Pi ) +
L (l) r j i
(30.3)
j ∈Ci \ j
Step 4
Calculate the sum of LLR: L(Q i ) = L(Pi ) +
L r ji
(30.4)
j∈Ci
If the convergence condition is met or the number of iterations reaches the threshold, the operation ends. If not, repeat from step 2. Then, calculate hard decision messages for all variable nodes. If L(Qi ) > 0, then the estimated bit value oi = 0, otherwise oi = 1.
30.2.2 Model-Driven Deep Learning Deep learning is a standard data-driven approach, which takes deep network as a black box and relies on a large amount of data to solve real problems. The modeldriven method starts from the priory of the target mechanism to form a learning cost function and then minimizes the cost function to solve the problem. In the field of communication, model-driven deep learning can build network architecture based on existing communication models. At present, this method has been successfully used in communication, including the improvement of channel decoding algorithm [6, 7].
254
Y. Tang et al.
30.3 Neural BP Structure In this chapter, we will introduce how to realize BP decoding in neural network and the detailed design information of each layer of the network. First, Fig. 30.1 gives the architecture of the proposed decoding network. The code word x of length k is encoded into the binary code word u of length N and then modulated by BPSK. The BPSK symbol will be passed through a channel with Gaussian noise. At the receiver, use the receiver signal y = s + n to calculate the LLR value and then feed it into the deep feedforward neural network. Then, we give a detailed signal flowchart on the receiver side of the decoding network structure, as shown in Fig. 30.2. The iterative decoding algorithm is expanded into a forward feedback network. Each hidden layer corresponds to a BP iteration process. The hidden layer contains two types of neurons: one is the CN neuron which calculates the extrinsic information in the CN layer: ⎛ L (l) (r ji ) = 2 tanh−1 ⎝
⎞ 1 (l−1) ⎠ qi j tanh L 2 \i
i ∈R j
(30.5)
Another one is VN neuron. It calculates the extrinsic information in the VN layer:
Gaussian noise
LDPC Encoder
BPSK
Input Layer
Hidden Layer
Output Layer
Fig. 30.1 Process diagram of the decoding algorithm for neural BP network
Fig. 30.2 Signal flowchart of model-driven deep learning BP decoding network
30 Modern-Driven Deep Learning for Belief Propagation LDPC Decoding
L (l) (qi j ) = L(Pi ) +
255
L (l) r j i
(30.6)
⎞ L (lmax ) r ji ⎠
(30.7)
j ∈Ci \ j
In the output layers, there is: ⎛ oi = σ ⎝ L(Pi ) +
j∈Ci
−1 We use σ = 1 + e−x as activate function in output neurons. It is easily to increase the iterations by increasing the CN layer and VN layer of the hidden layer. Besides, we choose cross entropy as loss function. It is defined as: L(P, o) = −
N 1 Pi log(oi ) + (1 − Pi ) log(1 − oi ) N i=1
(30.8)
30.4 Simulation Result and Analysis In this section, we analyze the performance of neuron BP LDPC decoding algorithm through numerical simulation to verify the effectiveness of the algorithm. To facilitate training, we use the Tensorflow framework [8]. The training data uses the code with length 567 and rate 3/4. The parity check matrix H was selected from [9]. Network training uses mini-batch gradient descent. There are 120 examples in each mini-batch. In addition, we applied the Adam optimization method for searching the optimal network parameters with learning rate equals 0.03. In order to test the robustness of the decoder, we choose to use varying SNR to generate training data. SNR equal to 0 dB means that the channel condition for data generation is poor, and many errors are generated in the data set, which will make the network unable to correctly learn the structure of the decoder, resulting in under-fitting of the network. SNR equal to 5 dB means that the channel conditions for data generation are very good, and there will be very little error data feedback input into the network, which will make the network generalization ability poor and cannot eliminate noise correctly. As shown in Fig. 30.3, compared with using 0 dB or 5 dB SNR alone to generate training data, the performance of the decoder will be better with data trained in multiple channels conditions of 0–5 dB. Figure 30.4 shows the relationship between BER performance and the number of network layers. It can be seen that in a certain network layer, the BER performance improves with the increase of the number of iterations. But the network has more layers, the complexity is higher. Therefore, after comprehensive consideration, a 44-layer network was finally selected for 20 iterations.
256
Fig. 30.3 BER performance under different SNR conditions
Fig. 30.4 BER performance under different number of network layers
Y. Tang et al.
30 Modern-Driven Deep Learning for Belief Propagation LDPC Decoding
257
30.5 Conclusion This article introduces the iterative unfolding process of BP decoding algorithm and then proposes an improvement scheme based on forward feedback. The improved scheme expands the iterative process of decoding into the neural network and uses the self-learning characteristics of the neural network to perform more reliable decoding. We believe that the network structure of the neural network can be used to increase the number of iterations of decoding to improve BER.
References 1. Gallager, R.: Low-density parity-check codes. IRE Trans. Inf. Theory 8(1), 21–28 (1962) 2. Mackay, D.J.C., Neal, R.M.: Near Shannon limit performance of low density parity check codes. Electron. Lett. 32(18), 1645–1646 (1996) 3. Nachmani, E., Be’ery, Y., Burshtein, D.: Learning to decode linear codes using deep learning. In: 2016 54thAnnual Allerton Conference on Communication, Control, and Computing (Allerton), pp. 341–346. IEEE, Monticello, IL, USA (2016) 4. Wang, Q., Wang, S., Fang, H.: A model-driven deep learning method for normalized min-sum LDPC decoding. In: 2020 IEEE International Conference on Communications Workshops (ICC Workshops), pp.1–6. IEEE, Dublin, Ireland, Ireland (2020) 5. Tanner, R.M.: A recursive approach to low complexity codes. IEEE Trans. Inf. Theory 27(5), 533–547 (1981) 6. He, H., Jin, S., Wen, C.-K., Gao, F., Li, G.Y., Xu, Z.: Model-driven deep learning for physical layer communications. IEEE Wirel. Commun. 26(5), 77–83 (2019) 7. Zhang, J., He, H.: Yang X: Model-driven deep learning based Turbo-MIMO receiver. In: 2020 IEEE 21st International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), pp. 1–5. IEEE, Atlanta, GA, USA (2020) 8. Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Cito, C., Corrado, G.S., Davis, A.: Tensorflow: Large-scale machine learning on heterogeneous systems. arXiv preprint arXiv:1603. 04467 (2016) 9. Channel Codes. www.uni-kl.de/channel-codes. Last accessed 2020/11/27
Chapter 31
Blind Recognition of Channel Coding Based on Deep Learning Shuying Zhang, Lin Zhou, Zizhuo Zhao, and Yiduo Tang
Abstract Blind recognition of channel coding means that the receiver can identify which channel coding type is used by the transmitter without any prior knowledge of channel coding. Traditional blind recognition algorithms for channel coding have high computational complexity and can only recognize specific coding type. In this paper, an algorithm to blindly identify channel coding types via deep learning is proposed, which can automatically extract the signal features and avoid complicated calculation. In particular, quasi-cyclic low-density parity-check (QC-LDPC) code and spatially coupled low-density parity-check (SC-LDPC) code are used as candidate codewords, and these two coding types are identified from noisy signals. Two kinds of recognizers based on long short-term memory (LSTM) network and convolutional neural network (CNN) are proposed, and the channel coding types are recognized through the classification process of neural network. Experimental results show that the proposed blind recognition method is able to realize high recognition accuracy even in poor channel conditions.
31.1 Introduction Channel coding is to resist noise interference by inserting check bits in the information sequence through certain rules. LDPC code has attracted wide attention because of its performance approaching Shannon limit. In a cooperative communication system [1], the receiver can know the channel coding type and parameters adopted by the sender through the control channel. The receiver can decode correctly according to the information. In the environment of non-cooperative communication [2], the receiver must identify which channel coding is used by the sender in order S. Zhang · L. Zhou · Z. Zhao (B) · Y. Tang Xiamen Key Laboratory of Mobile Multimedia Communications, National Huaqiao University, Xiamen 361021, Fujian, China e-mail: [email protected] L. Zhou · Z. Zhao State Key Laboratory of Integrated Services Networks, Xidian University, Xi’an 710071, Shaanxi, China © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_31
259
260
S. Zhang et al.
to decode the received signal correctly, and this identification is usually based on the unknown information of the sender. In this paper, we propose two recognizers based on LSTM and CNN, which can blindly identify the channel coding type through the classification process of deep learning. In our method, the recognizer first goes through a training stage, in which, the two recognizers are trained using data with known labels. Since the training dataset is a small part of all possible codeword datasets, it may cause the recognizer to be unable to correctly recognize the coding sequence that are not exist in the training dataset. However, through a large number of experiments, we prove that using the proposed deep neural network structures, such as LSTM and CNN, we can train a recognizer with good performance without any knowledge of channel coding.
31.2 Problem Formulation It is assumed that the information sequence before channel coding is b = [b1 , b2 , . . . , b K ], the sequence after channel coding is c = [c1 , c2 , . . . , c N ], usually N > K , and the sequence after modulation is s = [s1 , s2 , . . . , s N ]. The sequence received by the receiver is r = [r1 , r2 , . . . , r N ]. Denote the candidate set of codes by = {C1 , C2 , . . . , Cn }, where n is the total number of channel coding types adopted by the sender. The receiver needs to identify which encoding type in the set is used by the sender according to the signal r and the set [3]. Another problem is that the receiver needs to judge whether C x is used by the sender according to the received sequence r and the known coding scheme C x , so as to decide whether r can be decoded correctly by C x . In this paper, we focus on solving the former.
31.3 Deep Learning Method In this section, we introduce the proposed recognizer based on LSTM and CNN, and describe their training process.
31.3.1 LSTM Recognizer Recurrent neural network (RNN) is good at dealing with problems related to time series [4]. There is a certain correlation between different bits of the code word after channel coding, so RNN can be considered to solve this problem. LSTM is a variant of RNN, which overcomes the problem that the original RNN is difficult to deal with long-term dependence. Therefore, we expect it to perform well in dealing with coding sequences with long dependencies.
31 Blind Recognition of Channel Coding …
261
Fig. 31.1 Structure of LSTM recognizer
The LSTM structure is shown in Fig. 31.1. There are two ‘state vector’ in the hidden layer of LSTM, state h and state c. State h is very sensitive to short-term input, and state c can preserve long-term state. Thanks to this structure, LSTM can deal with long-time series problems well. In this paper, we first divide the observed sequence r into f segments, then the length of each section is d = N / f , where N is the length of the sequence r . Note that it is necessary to choose an appropriate f to ensure that d is an integer. Then, we input the whole observed sequence into LSTM network according to f time steps. At the kth time, LSTM received the kth sequence δk as input, the instant state of hidden layer hk = g(hk−1 , δk ) contains the current input and all the historical information of the past moment. The unit state ck = σ (ck−1 , hk ), which determines how much of the previous unit state ck−1 and the current hidden layer state hk remain to the current time ck . g(·) and σ (·) are nonlinear functions, which depend on the structure of LSTM. At last, the state c f is connected to the softmax activation function to identify the channel coding type and outputs the estimated probability p.
31.3.2 CNN Recognizer Compared with the traditional deep neural network, CNN because of its advantages of local connection and weight sharing, which greatly improves the speed of model training. CNN is usually use to process topological data with grid structure [5]. Convolution layer and pooling layer are typical structures of CNN [6], as shown in Fig. 31.2. Convolution layer is the most important part of CNN. After data input, the filter traverses the data with a specific step size, that is, performs convolution operation. In this paper, the maximum pooling algorithm is adopted, that is, the maximum value in the corresponding area is extracted with a specific step size to form a new feature map. We first convert one-dimensional coding sequence into two-dimensional, and then input it into CNN, as shown in Fig. 31.3. Two-dimensional data passes through convolution layer, pooling layer and full connection layer in turn, after which a softmax function is connected and, finally, outputs the estimated probability p.
262
S. Zhang et al.
Fig. 31.2 a Convolution layer with filter size of 3 × 3 and step size of 1; b Maximum pooling layer with filter size of 2 × 2 and step size of 2
Fig. 31.3 Structure of CNN recognizer
31.3.3 Training Process Training set is a dataset of known labels, in other words, a sequence of known coding types. p is one-hot coding representation of all coding type:
pˆ = (C1 , C2 , . . . Cn )
(31.1)
where n is the number of encoding types. If Ci (i = 1, 2, . . . , n) is the coding type used by the sender, Ci = 1, and other C j = 0( j = i). In the training process, we use the sequence of known coding types to find the appropriate deep neural network parameter θ, which makes: p) θ ∗ = argmin L(p,O θ
(31.2)
where L is the loss function [7]. The final ideal result is that the loss value is infinitely close to 0 through training.
31 Blind Recognition of Channel Coding …
263
31.4 Experimental Results In this section, we give some experiments to evaluate the performance of the recognizer based on deep learning. In this paper, TensorFlow deep learning framework is used, and GPU is used to accelerate training. In order to use the deep learning model for training, it is necessary to generate a large amount of sample data. In this paper, SC-LDPC code and QC-LDPC code with a rate of 1/2 and lengths of 36, 72, 144, 288, 400 and 512, respectively, are generated by MATLAB, with 100,000 frames each. Then, the encoded sequence is modulated by binary phase shift keying (BPSK) and then transmitted in additive white Gaussian noise (AWGN) channel. The signal-to-noise ratio (SNR) is 0–8 dB, and the demodulated sequence constitutes the dataset. In the sample dataset, the ratio distribution of training set and test set is 8:2. The loss function adopts cross entropy loss function, and the gradient descent is Adam adaptive learning rate optimization algorithm. Figure 31.4 depicts the changing trend of the accuracy of the two recognizers with the sequence length from 36 to 512 when SNR = 5 dB. It can be seen from the figure that the accuracy of the two LDPC codes increases with the increase of sequence length, and the performance of CNN recognizer is slightly higher than that of LSTM recognizer. Generally speaking, the recognition accuracy of the two recognizers is very high. Even when the sequence length is short, the two curves are always higher than 93%. The recognition accuracy is close to 100% when the sequence length is greater than 300. These experimental results show that our recognizer is excellent in recognizing different types of LDPC codes. Figure 31.5 depicts the changing trend of the accuracy of the two recognizers with the SNR from 0 to 8 dB when sample length of 512. Obviously, the coding classification performance of the two algorithms increases with the increase of SNR. The performance of the two algorithms based on deep learning is almost the same. Fig. 31.4 Identification accuracy of SC-LDPC and QC-LDPC with the sequence length from 36 to 512 when SNR = 5 dB
264
S. Zhang et al.
Fig. 31.5 Identification accuracy of SC-LDPC and QC-LDPC with the SNR from 0 to 8 dB when sample length of 512
When the SNR is 0 dB, the recognition accuracy is greater than 80%, and when the SNR is greater than 5 dB, the recognition accuracy is close to 100%. The reason for this phenomenon may be that under the condition of low SNR, the neural network learns too many noise patterns in the training stage, which leads to over-fitting. In any case, the results show that our algorithm can work well in the low SNR region.
31.5 Conclusion In this paper, we consider using deep learning method to realize blind recognition of channel coding. This method can effectively avoid the complex operation of extracting data features manually and save a lot of labor costs. Meanwhile, the experimental results show that this method can still effectively identify the coding type under the condition of more complex channel environment. In the future work, we will extend our method to identify more LDPC codes with different code rates. In addition, machine learning methods and other channel coding such as convolutional codes, turbo codes, and polar codes will also be considered.
References 1. Haykin, S.: Cognitive radio: brain-empowered wireless communications. IEEE J. Sel. Areas Commun. 23(2), 201–220 (2005) 2. Ali, A., Hamouda, W.: Advances on spectrum sensing for cognitive radio networks: theory and applications. IEEE Commun. Surv. Tutor. 99, 1277–1304 (2017) 3. Yu, P., Peng, H., Li, J.: On blind recognition of channel codes within a candidate set. IEEE Commun. Lett. 20(4), 736–739 (2016)
31 Blind Recognition of Channel Coding …
265
4. Peng, S.L., Jiang, H.Y., Wang, H.X., Alwageed, H.: Modulation classification based on signal constellation diagrams and deep learning. IEEE Trans. Neural Netw. Learn. Syst. 99, 718–727 (2018) 5. Ni, Y., Peng, S., Zhou, L., Yang, X.: Blind identification of LDPC code based on deep learning. In: 2019 6th International Conference on Dependable Systems and Their Applications (DSA), Harbin, China, pp. 460–464 (2020) 6. Qin, X., Peng, S., Yang, X., Yao, Y.: Deep learning based channel code recognition using TextCNN. In: 2019 IEEE International Symposium on Dynamic Spectrum Access Networks (DySPAN), Newark, NJ, USA, pp. 1–5 (2019) 7. Shen, B., Wu, H., Huang, C.: Blind recognition of channel codes via deep learning. In: 2019 IEEE Global Conference on Signal and Information Processing (GlobalSIP), Ottawa, Canada, pp. 1–5 (2019)
Chapter 32
Influence and Optimization of Levenberg–Marquardt Algorithm on Dielectric Loss Measurement in Noisy Environment Chao Yu, Ronghui Hu, Xiaoxiao Xu, and Suaisuai Zhao Abstract In the calculation of medium loss angle, Levenberg–Marquardt algorithm (LM algorithm for short) can improve the high-order sinusoidal fitting method in the case of no noise, and the calculation amount is large and the speed is relatively slow. However, white noise and quantization noise exist in the actual measurement of dielectric loss angle. In order to extend LM algorithm to the actual measurement of medium loss angle, this paper first studies the influence of white noise on LM algorithm in the number of iterations and calculation accuracy and then studies the change of LM algorithm after signal noise reduction through FIR narrowband bandpass filter, as well as the change of iteration times and calculation accuracy. The research shows that noise level is positively correlated with the iteration times of LM algorithm and negatively correlated with the accuracy. After noise reduction, LM algorithm can be greatly simplified, the number of iterations is reduced, and the accuracy is also improved.
32.1 Introduction The dielectric loss Angle δ is the complementary angle of the phase difference between voltage and current under sinusoidal AC voltage. The tanδ is called the tangent of medium loss, which is an important index reflecting the insulation performance of high voltage electrical equipment [1]. The change of dielectric loss Angle can reflect the insulation defects such as damp exposure, deterioration or gas discharge in the insulation, so the detection of dielectric loss of electrical equipment is of great value to the safe operation of power system [2, 3]. Dielectric loss Angle is a small value, so it is greatly affected by voltage measurement. However, the voltage measurement value is easy to be distorted by frequency offset and harmonic interference [4]. In the calculation of medium loss, one of the most basic algorithms is the least square method [5], but this algorithm only considers the fundamental wave, treats the higher harmonics as noise, and thinks that the power C. Yu (B) · R. Hu · X. Xu · S. Zhao Department of Electrical Engineering, Shanghai Dianji University, SDJU, Shanghai, China © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_32
267
268
C. Yu et al.
grid frequency remains unchanged at 50 Hz. Then the improved algorithm is higherorder sine fitting algorithm [6]. The variation of high harmonics and power grid frequency is considered, but the calculated equation is a high-order nonlinear equation, which requires the use of Newton iterative algorithm, and the initial value of iteration is obtained by the discrete Fourier algorithm. Since the Hessian matrix in Newton iterative algorithm is a second-order matrix, calculation and code implementation are troublesome, and iteration takes up a lot of computing resources, so it will take a long time to get a more accurate value. Later, the improved algorithm uses the Levenberg–Marquardt algorithm (abbreviation is LM) to iterate [7], and the algorithm uses Jacobian matrix and the identity matrix with parameters to approximate the Hessian matrix. In reference 6, under the condition of no noise, the LM algorithm is superior to the Newton iteration method in terms of algorithm simplicity and final fitting results and speed. But in practical engineering, there will be white noise, impulse noise and quantization noise [8, 9], and white noise is the most common noise that has the greatest impact on the results. Therefore, in order to extend LM algorithm to the actual measurement of dielectric loss, the change of the accuracy and iteration times of LM algorithm caused by white noise, as well as the improvement of the algorithm in this case, is the direction considered in this paper.
32.2 High-Order Sinusoidal Fitting Method Is Used to Calculate the Dielectric Loss The fast Fourier algorithm can obtain the amplitude and phase of each frequency in the case of quasi-synchronous sampling [10], but the power grid frequency is not a fixed value and fluctuates to a certain extent. Therefore, sampling cannot be carried out synchronously, and spectrum aliasing will occur. Although the higherorder sinusoidal fitting is also based on the Principle of Fourier transform, it takes frequency as the unknown quantity and uses the method of harmonic weighting to fit the voltage and current signal. Decompose the measured voltage and current signal into. In the Fig. 32.1, the measured waveform is decomposed into DC component, fundamental wave and odd harmonic component (even harmonic amplitude is very small and can be ignored), and the fundamental wave is decomposed into sine and cosine components again for the convenience of calculation. The power grid
Fig. 32.1 Original waveform decomposition
32 Influence and Optimization of Levenberg–Marquardt Algorithm …
269
frequency is a fluctuating value, but it can be assumed that the power grid frequency does not change in a very short time. This paper assumes that the power grid frequency remains unchanged in a power frequency cycle. The current and voltage signals can be represented by the following mathematical model: I (tn ) = A0 +
N
(A K 1 sin((2k − 1)2π f tn ) + A K 2 cos((2k − 1)2π f tn ))
(32.1)
k=1
U (tn ) = B0 +
N
(B K 1 sin((2k − 1)2π f tn ) + b K 2 cos((2k − 1)2π f tn ))
(32.2)
k=1
In this equation, A0 , B0 is the DC component; Ak1 , Ak2 , Ak3 , and Ak4 are the amplitude of the sine and cosine components of the (2k − 1) harmonic, respectively, t n is the nth time point. The two mathematical models are used to fit the actual voltage and current signals. The fitting error can be expressed as: ei (n) = I (tn ) − i(tn )
(32.3)
eu (n) = U (tn ) − u(tn )
(32.4)
The squared sum representation of the error is as follows: Ei =
M−1
(I (tn ) − i(tn ))2
(32.5)
(U (tn ) − u(tn ))2
(32.6)
n=0
Eu =
M−1 n=0
In this equation, i(t n ) and u(t n ) are the current and voltage series obtained in practice. In order to minimize the sum of squares of errors, an appropriate nonlinear optimization algorithm can be used to achieve this. After each parameter is determined, the tangent of the medium loss Angle can be expressed as: tan δ =
A12 B12 + A11 B11 A11 B12 − A12 B11
(32.7)
In this equation, A11 , A12 , B11 , and B12 are the amplitude of sine and cosine components of the fundamental wave in the voltage and current signal.
270
C. Yu et al.
32.3 Levenberg–Marquardt Algorithm The LM algorithm is a numerical solution that provides nonlinear minimization. By modifying parameters during execution, the proposed algorithm achieves the advantages of both Gauss–Newton algorithm and gradient descent algorithm and improves the shortcomings of both algorithms. When LM algorithm is used, the recursive equation is: X(m + 1) = X(m) − H i −1 J i (m)T ei (m)
(32.8)
Y (m + 1) = Y (m) − H u −1 J u (m)T eu (m)
(32.9)
In this equation, X and Y are variables of nonlinear functions of multiple variables, X(m) and Y (m) are the results obtained by the mth iteration, X(m + 1) and Y (m + 1) are the result of the No (m + 1) iteration, H i and H u is the Hessian matrix, J i and J u is the Jacobian matrix, ei (m) and eu (m) are the corresponding error column vector, and m is the number of iterations. To reduce repetition, take current as an example to expand the recursive equation, where the expression of H matrix is: H i (m) ≈ J i (m)T J i (m) + λI
(32.10)
In the equation, I is the 2N + 2 orders identity matrix, λ is an adjustable parameter, and J i (n) is Jacobian matrix. This matrix is obtained by taking the first-order derivative from I. The first-order derivative at the nth point can be expressed as: J in =
∂ I (tn ) ∂A
(32.11)
In this equation, A is the 2 + 2N dimensional variable vector, A = {A0 ,f . A11 , A12 , …, Ak1 , Ak2 }, Ak1 , and Ak2 are the amplitude of the (2k − 1) harmonic sine and cosine component. The partial derivative of each variable is obtained as follows: Jin,1 =
∂ I (tn ) =0 ∂ A0
(32.12)
∂ I (tn ) = = (Ak1 (2k − 1)2π tn cos((2k − 1)2π f tn ) − Ak2 (2k − 1) ∂f k=1 n
Jin,2
2π tn sin((2k − 1)2π f tn )) Jin,2k+1 =
∂ I (tn ) = sin((2k − 1)2π f tn ) ∂ Ak1
(32.13) (32.14)
32 Influence and Optimization of Levenberg–Marquardt Algorithm …
271
∂ I (tn ) = cos((2k − 1)2π f tn ) ∂ Ak1
(32.15)
Jin,2k+2 =
If the data sampling is out of sync, the results obtained by directly calculating the signal parameters using FFT are not accurate. However, due to the small fluctuation of the power grid frequency, it is expected that the obtained results are not different from the actual signal parameters. Therefore, the initial value of iteration except the frequency (when m = 0) can be obtained by the current signal changing through fast Fourier, and the frequency is set as 50 by default. Taking these values as initial values can reduce the number of iterations and speed up the calculation. The flowchart of high-order sinusoidal fitting medium loss Angle based on LM algorithm is as follows (Fig. 32.2). To analyze the influence of white noise on LM algorithm, the frequency was fixed at 50 Hz, and only five or less harmonics were considered. The amplitude parameters Fig. 32.2 LM algorithm flow chart
272
C. Yu et al.
Fig. 32.3 Fundamental phase diagram of voltage and current under noise
of voltage and current signal are consistent with literature 6. The phase angles of the fundamental current wave, the third harmonic wave, and the fifth harmonic wave are 30°, 150°, and 75°, respectively. The phase angles of the fundamental, the third, and the fifth harmonics of the voltage are 60, 148, and 78°, respectively. Is the voltage and current plus two normal distribution random noises with the same SNR and the SNR range is: 5–55. The sampling frequency is 10,000, and 200 sampling points are taken for calculation. The results of the fundamental wave phase Angle of the voltage and current calculated are as follows (Fig. 32.3). The detailed value of the fundamental wave phase and the number of iterations are as follows: Table 32.1 shows that in LM algorithm, the amount of noise is positively correlated with the number of iterations and negatively correlated with the accuracy. Moreover, even when the noise is very small, it has a great impact on the accuracy of the final result. Therefore, in the actual measurement of medium loss, the sampling sequence of voltage and current needs to be digitally filtered first. Table 32.1 Fitting results of LM algorithm with different SNR Signal to noise ratio
Fundamental wave phase angle
The number of iterations
Voltage
Current
Voltage
Current
60.0000
30.0000
0
0
5
49.0255
26.2268
6
6
15
59.9795
27.7562
3
4
25
59.2745
31.1238
3
3
35
60.1334
29.9385
2
2
45
59.8481
30.0352
2
2
55
59.9899
30.073
2
2
No noise
32 Influence and Optimization of Levenberg–Marquardt Algorithm …
273
32.4 FIR Narrowband Bandpass Filter In the online measurement of dielectric loss, the required frequency is the frequency near the power frequency. If there is a narrowband-pass filter of 49–51 Hz, the high-order harmonics and most of the noise can be removed. In the calculation, the high-order harmonics need not be taken into account, and the reduction of noise can also reduce the number of iterations. The calculation of dielectric loss Angle is to find the complementary Angle of voltage and current phase difference, and the filter must be strictly linear phase to ensure that the calculated results will not generate new errors due to the filter. FIR filter is a full zero point system, that is, the Z transfer function at the Z complex plane poles are all at Z = 0, which makes the FIR filter have very good linear phase. FIR filter design methods have window function design method and frequency sampling design method. The frequency sampling design method is designed from the frequency domain, which is especially suitable for narrowband bandpass filters. Therefore, this paper uses the frequency sampling design method to design narrowband bandpass filters of 49–51 Hz. In the frequency range, suppose there is a target frequency response H d (ejw ), which is a continuous function about w, and now sample N points at equal intervals on the unit circle of ejw = 1, and get H d (k), and this is the actual frequency response: Hd (k) = Hd e jω |ωk =2πk/N = Hd e j2πk/N
(32.16)
H d (k) is transport by the inverse Fourier transform to the time domain, and the required impulse response is obtained: h(n) =
N −1 j2πk 1 Hd (k)e N , n = 0, 1, 2, . . . , N − 1 N k=0
(32.17)
When giving the required frequency response H d (k) in the time domain, generally only the value of 0 ~ π will be given, so the value of π ~ 2π needs to be filled up according to the linear phase condition of FIR filter before doing the Inverse Fourier transform. The ideal frequency response is defined as: H e jω = H (ω)e jθ(ω)
(32.18)
This expression represents the multiplication of amplitude response and phase response. To satisfy the linear phase condition, it can be expressed as:
N −1 H (N − k) = H (k) , k = 1, 2, . . . , N −1 θ (N − k) = N π k 2
(32.19)
274
C. Yu et al.
Fig. 32.4 Filter effect chart
In this equation, θ (k) = − NN−1 π k(k) = −(N − 1)kπ /N, when N is even, H(N/ 2) = 0. The order of FIR filter is large enough to make the filtered waveform smooth enough, and the filter has better linear phase when the order is odd, so the selected order is close to 599 of three periods, that is, 599 sampling points on the unit circle. The corresponding sampling points of 49–51 Hz are retained, and the remaining sampling points are set to zero. A standard sine wave with frequency of 50 Hz and a 50 Hz sine wave with noise are put into the filter, and the results are as follows (Fig. 32.4). As can be seen in the figure, this filter will not distort the original waveform, and the filtering effect can also meet the requirements. After filtering, the first part of waveform is different from the original waveform. This is because the FIR filter carries out convolution operation with signals in the form of window. Only after the whole window enters can the waveform be completely restored. Since the order of this filter is 599, the original waveform is completely restored from the 600th point. After calculation, the phase of the filtered waveform is shifted back by 178.2 degrees compared with the original waveform. In this paper, the filtered phase results are obtained after phase compensation. The simulation data of voltage and current is the same as the previous one, but 1000 sampling points are taken to enter the filter. After filtering, take 200 points from the 600th to 799 and enter into LM algorithm calculation. Because this digital filter can filter out all the higher harmonics, the variables are reduced to only A0, A11, A12, and F.
32 Influence and Optimization of Levenberg–Marquardt Algorithm …
275
Fig. 32.5 Fundamental phase diagram of voltage and current after filtering
The voltage and current signal are the same as the previous section. The phase Angle of the fundamental wave of the voltage and current calculated is as follows (Fig. 32.5). The detailed value of the fundamental wave phase and the number of iterations are as follows. The comparison between Tables 32.1 and 32.2 shows that the FIR narrowband bandpass filter can reduce the iteration times of LM algorithm when the noise is large and improves the accuracy of results when the noise is small. When the SNR is 55 and the frequency fluctuates between 49.5–50.5, with a filter, the results are as follows. When the SNR is 55, the frequency fluctuates between 49.5–50.5, and there is no filter, and the results are as follows. It can be seen from Tables 32.3 and 32.4 that, during the non-integral period sampling, the phase of the obtained signal will deviate from the actual value due to the fact that the point is fixed from the 600th point after filtering. However, the voltage and current offsets are basically the same, so the effect on the results can be ignored. So when noise and frequency fluctuations are present, and FIR bandpass filters can also improve the accuracy of the final results. Table 32.2 Fitting results of LM algorithm with different SNR after denoising Signal to noise ratio
Fundamental wave phase angle
The number of iterations
Voltage
Current
Voltage
Current
There is no noise
60.0000
30.0000
0
0
5
54.4482
29.6957
3
3
15
60.4249
30.6393
3
3
25
59.5213
29.8220
3
3
35
60.0593
30.1198
1
1
45
59.9796
30.0116
1
1
55
60.0124
30.0047
1
1
276
C. Yu et al.
Table 32.3 Fitting results of LM algorithm with different frequencies after noise reduction Frequency
Fundamental wave phase angle
The number of iterations
Voltage
Current
Voltage
Current
49.5
54.5787
24.5723
3
3
49.7
56.7539
26.7527
3
3
50
59.9991
30.0016
1
1
50.2
62.1674
32.1630
3
3
50.5
65.4133
35.4185
3
3
Table 32.4 Fitting results of LM algorithm at different frequencies Frequency
Fundamental wave phase angle
The number of iterations
Voltage
Current
Voltage
Current
49.5
60.0266
29.9899
3
3
49.7
59.9797
29.9766
3
3
50
59.9965
30.0143
2
2
50.2
59.9819
30.0212
3
3
50.5
60.0053
29.9793
3
3
32.5 Conclusion The high-order sinusoidal fitting method can obtain the intermediate loss value in the case of unknown grid frequency, but LM algorithm is needed to solve the problem. However, in the actual measurement of dielectric loss, LM algorithm has a great error due to the presence of white noise, and the number of iterations also increases significantly. Therefore, for LM algorithm, noise reduction before calculation is necessary. FIR narrowband bandpass filter can filter out most of the noise and higher harmonics, greatly simplifying LM algorithm, algorithm iteration times in the noise more significantly reduced, and accuracy is also significantly improved. In addition, FIR filter can also improve the accuracy of results when frequency fluctuation and noise exist simultaneously.
References 1. Liao, J.Q., Liang, G.X.: A method for measuring dielectric loss combining complex trigonometric function and Hilbert transform. Insul. Mater. 45(3), 69–72 (2012) 2. Tian, C.F.: Correct measurement of capacitance capacity and dielectric loss of transformer capacitive bushing. Electr. Autom. Equip. 30(4), 147–149 (2010) 3. Peng, X., Chen, H., Xia, G.L. et al.: Field high voltage dielectric loss measurement of shunt capacitors with 500 kV circuit breakers. High Voltage Technol.37(10), 2378–2384 (2011)
32 Influence and Optimization of Levenberg–Marquardt Algorithm …
277
4. Zhang, W., Lv, Z.C., Deng, Y.R.: Application of fundamental wave frequency correction method in measuring dielectric loss. Insul. Mater. 4, 77–80 (2016) 5. Zhao, X.S., Tan, K.X., Zhu, D.H., et al.: Digital measurement of dielectric loss angle. J. China Univ. Sci. Technol. (Nat. Sci.) 36(9), 51–56 (1996) 6. Wang, W.L., Li, F.Q., Tan, K.X.: High order sinusoidal fitting method for measuring dielectric loss Angle. J. Tsinghua Univ. (Nat. Sci. Ed.) 41(9), 5–8 (2001) 7. Xu, Z.N., Feng, G.Y., Lu, F.C., et al.: Comparison between Newton method and LevenbergMarquardt algorithm and its application in mesoloss Angle measurement. J. North China Electr. Power Univ. 33(1), 15–19 (2006) 8. Chen, L.A.: Modified Levenberg-Marquardt method with line search for nonlinear equations. SCI Comput. Optim. Appl. 3, 753–779 (2016) 9. Xu, Z.N.: Influence of noise on the tangent calculation results of dielectric loss Angle. Electr. Power Autom. Equip. 33(11), 69–74 (2013) 10. Wang, M., Vandermaar, A.J., Srivastava, K.d.: Review of condition assessment of power insulation in service. IEEE Electr. Insul. Mag. 18(6), 12–25 (2002)
Chapter 33
Research on Artificial Intelligence Optimization Based on Genetic Algorithm Jingyi Liu
Abstract As an important force for national construction and development, science and technology are also a major factor affecting global economic development. Especially in the rapid development of artificial intelligence technology, genetic algorithm, as one of its optimization methods, has played a positive role in the field of innovation and technological innovation, and has changed people’s life and work in a certain sense. Therefore, on the basis of understanding artificial intelligence and genetic algorithm, this paper further studies how to carry out artificial intelligence optimization based on genetic algorithm.
33.1 Artificial Intelligence Artificial intelligence, also known as artificial intelligence, was first proposed by Hugo de Gareth, an American scientist. It refers to an artificial system that can sense, understand, make decisions and execute decisions as well as humans. Artificial intelligence can be understood in two aspects: artificial intelligence is human processing and manufacturing, while intelligence includes consciousness, thinking and self. In today’s social environment, the realization of intelligence requires three elements: first, big data; second, the Internet; third, cloud computing [1].
33.2 Genetic Algorithm To put it simply, genetic algorithm refers to the computational model constructed by means of genetics and natural selection in Darwinian biological evolution, which can search for the optimal solution. Generally speaking, the steps of genetic algorithm are mainly divided into the following: first, estimate the adaptation degree of each chromosome; second, choose two individuals from a population and treat them as J. Liu (B) Qingdao University of Science and Technology, Qingdao, China e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_33
279
280
J. Liu
parents. Third, obtain the parents’ chromosomes and obtain the offspring in the cross experiment; fourth, the chromosome variation of the progeny was treated. Fifth, repeat steps 2, 3, and 4 until a new species is created. Among them, the basic process box is shown in Fig. 33.1, and the pseudo-code of the simple operation process of genetic algorithm is shown in Fig. 33.2. In the cycle operation, researchers can find satisfactory solutions to various problems, and the technical terms involved are shown in Table 33.1 [2–4]. Fig. 33.1 Basic flow diagram
Fig. 33.2 Simple operation flow pseudo-code
33 Research on Artificial Intelligence Optimization …
281
Table 33.1 Description of biogenetic terms in genetic algorithm The term
explain
1 Chromosome
Also known as genotype individuals, having a certain number of individuals forms a population, and the number of individuals in a population is the size of the population
2 Gene
Used to represent individual characteristics, such as a string S = 1011, where 1, 0, 1, and 1 are genes
3 Gene loci
Represents the position of a gene in the string, also referred to as the gene position, for example, the position of S = 1011, 0 is 2
4 The eigenvalue If integers are represented by strings, the eigen value of a gene has the same weight as the binary number 5 Fitness
It refers to the degree to which an individual ADAPTS to the environment. To demonstrate the fitness of a chromosome, a function of measure is introduced for each chromosome in question, also known as the fitness function
33.3 Research on Artificial Intelligence Optimization Based on Genetic Algorithm 33.3.1 Robocode In July 2001, the Robocode proposed by IBM of the United States was a tank robot combat simulation engine, which had the following advantages: first, it was based on Java development and had object-oriented concepts such as inheritance; second, it can quickly test the genetic algorithm designed by itself through competition, which is interesting and competitive to some extent. Third, the community has 4000 robots to choose from, and can simulate the real genetic environment, more have a lot of code for the development of reference; fourth, Robocode itself is open source software, and designers can directly add new genetic features to promote heredity [5–7].
33.3.2 Optimization of Default Strategies There are many genetic algorithms that can be selected for robot optimization, such as genetic movement, preset strategy, etc. The following is to analyze how to scientifically optimize Robocode by taking preset strategies as an example. Since the basic operations of such robots are scanning, moving, shooting, etc., the specific code is as follows: radar scanning strategy is radar-strategy (RS), moving strategy is move-strategy (MS), bullet energy is bullet-power-strategy (BPS), etc.
282
J. Liu
Typically, robots like Robocode are programmed with four strategies: move, aim, radar, shoot, and if something happens, a behavioral strategy for that event is triggered. Because each strategy include different forms of processing, and can be triggered by genetic algorithm (GA), there may be about 4224 kinds of possibilities, so involved in mobile function also has A lot of kinds, like set A head and ahead can let robot continues to move forward A distance, and the set Back and the Back is backward, Part of the code as shown below, the following: Random: switch(Math.randlom0*2)[ case 0;sen TurnRight(Math.randomy*90):break; casel: set’TurnLeft(Math.random0*90); break;} exeeute(; Linear: aheacl(200):setBaek(200);Aheacl(200): setBaek(200); Circular: setTurnRighn 1000; setMax Nelocity(4):ahead( 1000): SETTurnRighn 100; SETMAX Nelcity(): be in the lead; (ake the) lead; in the highest fight; keep ah. ead; on top (1000): By using genetic algorithm to adjust the Robocode motion position, we can start from two aspects: first, master the position of the robot, make clear our own speed, and know the position, relative angle and speed of the opponent; second, master the robot’s motion position and know the opponent’s position, relative angle and speed. Second, the orientation, bullet position and size parameters of all robots on the field. Input the above information into the system, which can help the robot to calculate the accurate angle and distance when it moves. It should be noted that the program code should meet the following requirements: first, use the function to get the specific value of fitness; Second, genetic algorithm is used to operate the robot when it moves. Third, in order to obtain the genetic offspring in the exercise, you must not manually enter it after the end [8, 9]. Because the genetic algorithm is to achieve the movement must choose two categories, one is GA, and the other is move pattern. The former is relatively simple, mainly taking values and defining groups, and then reading and writing files based on the definition and data manipulation. The common data include the probability of crossover and mutation, population size and the weight coefficient of the equation. The latter can complete the movement according to cross variation, and the output content will be stored in the vector function. Since this function contains a set of real numbers, it can be used not only to calculate the x and y axes, but also to compute precise values such as distances and angles. In move pattern, for example, Grad, Input, and Rang represent the order, the input number and the weighted range of the function, respectively. During cross movement, a new feature code will appear in each operation, which is an array of variables between 0 and 1, and operations can be carried out according to the
33 Research on Artificial Intelligence Optimization …
283
weighted value in the function vector. As the operation of this position is relatively simple, any value within the weighted range is randomly selected to replace the array of variables and stored in the movement mode. The specific variation code is shown in below: protected void mutate0{ Reights X[inty (Math.random 0"weightsX.length)] = Math.randomk); range*2range; weightsY(int) (Math.random (*weight.Xlenghy] = Math.randlomJ"range2- range; After the completion of the above genetic algorithm optimization work, if the Robocode does not store the values in the working state, then the system will randomly acquire the new movement mode according to the input strategy, otherwise the system will automatically load and store the values. In the initial stage, the corresponding fitness will be proposed for each mobile mode, and the system will accurately number and implement the following steps: first, the mobile mode will be graded according to the specific value of fitness; second, elite, crossover and mutation operation should be carried out according to the requirements. Third, store the weighted value; fourth, redo the genetic algorithm. In the continuous adjustment of the system, researchers can get the best adaptive value, and thus, propose a more valuable optimization scheme.
33.4 Conclusion To sum up, compared with the previous optimization forms, genetic algorithm consumes less time, has high accuracy, is simple to operate, has strong expansibility and inspiration, and can integrate with other algorithms faster. Therefore, in the artificial intelligence in the process of rapid development, in the face of more and more high-technical requirements, while GA has made certain achievements, but in order to better solve the problem of artificial intelligence optimization, dealing with intelligent perception, and cognitive function in the system, must want to continue to strengthen research efforts, and combine with practice case analysis, only in this way can quickly achieve expected goals. And in future development, I believe that through our continuous exploration, we will be able to make new breakthroughs in the manufacture and application of artificial intelligence technology.
284
J. Liu
References 1. Lin, M.T., Zhao, Y.P.: Artificial intelligence-empowered resource management for future wireless communications: a survey. China Commun. 17(03), 58–77 (2020) 2. Patle, B.K., Ganesh Babu, L., Pandey, A., Parhi, D.R.K., Jagadeesh, A.: A review: on path planning strategies for navigation of mobile robot. Def. Technol. 15(04), 582–606 (2019) 3. Lin, C., Li, S., Wu, K., Liu, T.: Intelligent assessment method of air-defense & anti-missile command model based on genetic algorithm optimization. In: IOP Conference Series: Materials Science and Engineering, vol. 435(1), pp. 66–71 (2018) 4. Zhao, X., Li, Y., Boonen, P.: Intelligent optimization algorithm of non-convex function based on genetic algorithm. J. Intell. Fuzzy Syst. 35(4), 15–159 (2018) 5. Azadeh, G., Hossein, B., Isa, E., Majid, M., Bahram, G., Saeed Reza, K.: Uncertainty analysis of intelligent model of hybrid genetic algorithm and particle swarm optimization with ANFIS to predict threshold bank profile shape based on digital laser approach sensing. Measurement 121 (2018) 6. Li, A., Zhao, W.Z., Li, S.M., Qiu, X.Y., Wang, X.B.: Research on the motion trajectory optimization method based on the improved genetic algorithm for an intelligent vehicle. Proc. Inst. Mech. Eng. 230(13) (2016) 7. Wang, X.R., Zhou, F., Wang, X.W., Guo, J.: Intelligent optimization of the structure of the large section highway tunnel based on improved immune genetic algorithm. Chem. Eng. Trans. (CET J.) 59 (2017) 8. Li, A.J., Zhao, W.Z., Li, S.M., Qiu, X.Y., Wang, X.B.: Research on the motion trajectory optimization method based on the improved genetic algorithm for an intelligent vehicle. Proc. Inst. Mech. Eng. Part D: J. Automob. Eng. 230(13), 99–105 (2016) 9. Zhu, M., Su, B., Ning, G.M.: Intelligent optimization of urban real-time traffic route based on DNA non-dominated sorting genetic algorithm. Appl. Mech. Mater. 3744 (2015)
Chapter 34
Multi-channel Scheduling for Intelligent Information Collection Xingchuan Bao, Jianhua Yang, Jian Yang, Lin Peng, and Min Xu
Abstract In order to effectively improve the service quality of intelligent information collection based on wireless sensor networks in power emergency scenarios, further enhance the real-time data in intelligent information collection and reduce delay, a priority-based multi-channel scheduling strategy is proposed. First, based on real-time channel link state information, a link routing algorithm based on minimum hop spanning tree (LB-MHST) is proposed to overcome radio frequency interference and ensure the service quality of the smart grid. Then, according to the different delay requirements of different data packets in intelligent information collection, data priority transmission is considered, which effectively improves the data transmission efficiency of the sensing node and further meets the Quality of Service (QoS) performance requirements in intelligent information collection. Experimental results show that compared to the minimum hop spanning tree (MHST) algorithm, especially in multi-channel delay-aware transmission, the proposed algorithm improves the delay performance of high-load traffic in the case of single channel, 8 and 16 channels, respectively, increased by 12%, 15.2%, and 18%.
34.1 Introduction In order to ensure the QoS [1] in smart grid applications, in terms of communication interference, the influence of radio frequency interference [2] is often overcome through multi-channel communication, which not only ensures multi-channel communication, but also improves the service quality in smart grid. In terms of latency, parallel transmission is adopted to improve the overall performance of the X. Bao (B) · L. Peng · M. Xu Global Energy Interconnection Research Institute, Nanjing 210003, Jiangsu, China X. Bao State Grid Laboratory of Electric Power Communication Network Technology, Nanjing 210003, China J. Yang · J. Yang State Grid ZheJiang Electric Power Co.,Ltd, Hangzhou 310000, Zhejiang, China © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_34
285
286
X. Bao et al.
network [3]. But with the development of smart grid, the grid application is becoming more and more high to the requirement of service quality, make the most of the communication technology (as shown in Table 34.1) cannot fully guarantee required in the application of grid service quality, which requires a reliable and efficient communication architecture for smart grid applications in future to provide satisfactory service [4]. However, different communication architectures have different communication delay requirements [5]. In addition, the network traffic load and data type of data communication between communication devices are constantly changing [6], making time a key factor for data collection of smart grid. Intelligent channel allocation strategy is the most effective way to counter channel interference. Musham et al. [7] proposed the use of a near-optimal algorithm for grid WMNs to mitigate the impact of channel interference in multi-channel transmission. Suryawanshi et al. [8] proposed to construct a multi-channel communication model against data loss caused by excessive channel interference load and recover data lost due to channel interference as far as possible. Jayalakshmi et al. [9] proposed a routing tree based on the channel information collection of network nodes for channel allocation, so as to combat interference more effectively. However, the sensor network QoS performance is not only affected by rf interference, but also by the network topology. In the multi-channel allocation strategies against channel interference, the delay requirements of different packets are also different [10]. To further satisfy the QoS requirement in the application of information intelligent acquisition and awareness, this paper proposes a link routing algorithm based on the link grid minimum hop spanning tree (LG-MHST) for data priority channel scheduling strategy on the basis of ensuring channel interference. In addition, the priority of data is fully considered during the channel allocation, which not only guarantees the good anti-interference performance of intelligent information collection, but also reduces the transmission delay of intelligent information collection, and further comprehensively meets the QoS requirements of intelligent information collection for data transmission. Table 34.1 Several existing QoS aware routing protocol in wireless sensor network Protocol
Delay
Throughput
Reliability
RTLD [10]
No
Yes
Yes
RPL [10]
Yes
Yes
Yes
BIOSARP [11]
Yes
No
No
RRR [11]
Yes
Yes
Yes
OQAP [11]
No
No
No
MRL-CC [12]
No
No
No
DARA [12]
Yes
No
Yes
34 Multi-channel Scheduling for Intelligent …
287
34.2 Communication Model of the Main Station for Intelligent Information Collection According to research [11], the above model can be used in a radio propagation environment with interference, such as a state-aware environment for intelligent information collection. The relevant parameters of the model are shown in Table 34.2. In this model, the path fading from the transmitter d is: P L d0 →d = P L(d0 ) + 10η log10
d + Xσ d0
(34.1)
among them. P L d0 →d P L(d0 ) η Xσ
is the path fading at a distance d from the transmitter, unit is dBm. As reference distance d0 Path fading. Path fading coefficient. is Gaussian random variable with mean zero and standard deviation σ.
This paper is based on the good anti-interference performance of the receiver’s channel scheduling (RBCA) [12] algorithm, combined with WSN’s own network characteristics, in the later time gap allocation. The data priority is considered in parallel transmission scheduling with multi-branch, and finally a scheduling algorithm that divides channels according to data priority is proposed, that is, a link routing algorithm based on the minimum hop spanning tree. The purpose of reducing the delay of intelligent information collection and transmission is achieved, which further satisfies the QoS requirements of intelligent information collection for data transmission in complex environments. The LB-MHST algorithm takes the sink node in the WSN as the root node, so that the link cost of each sub tree connected to the sink node is less than the present value. Then, the channel scheduling algorithm based on data priority is shown in Fig. 34.1. Build a routing tree, then use the multi-channel scheduling algorithm based on time division multiple access (TDMA) to calculate the minimum scheduling length, and implement channel allocation according to the radio frequency (RF) status between nodes. That is, if the SINR between nodes exceeds a predetermined threshold, channels can be allocated to these nodes. Table 34.2 Logarithmic normal shadow model parameters in smart grid environment
Transmission environment
500 kV ground transmission (no unit)
Path fading (η)
2.42
Shadow variance (X σ )
3.12
288
X. Bao et al.
Fig. 34.1 Data priority channel scheduling algorithm
34.3 Algorithm First, based on the real-time channel link state information, based on the RBCA algorithm, an improved channel allocation algorithm is proposed. As shown in algorithm, first find the active parent node in the routing tree, and then assign channels to it. According to the delay requirements of information intelligent data collection, the node data packets that need to be forwarded are divided into three levels. According to different data priorities, a data transmission algorithm is proposed. Combining the above two algorithms, the channel allocation and data transmission based on data priority are finally achieved to improve the data transmission efficiency of the sensing nodes in the intelligent information collection. In algorithm, parents represent the parent node set, interfp represents the interfered parent node, sm represents the STN ratio matrix, and the threshold refers to the interference signal-to-noise ratio threshold. As shown in Fig. 34.2, after initializing the parent node, it is necessary to traverse all the parent nodes, compare the signalto-noise ratio thresholds one by one, and then traverse the nodes for classification. On the setting of the new parent node in the flowchart shown in Fig. 34.3, according to the example scheduling shown in Fig. 34.3, assuming that communication node
34 Multi-channel Scheduling for Intelligent …
Fig. 34.2 Flow chart of channel scheduling Fig. 34.3 Channel scheduling
289
290
X. Bao et al.
A is the parent node, and nodes B and C are within the communication radius of node A, it is calculated when A sends data. The signal-to-noise ratio values between A, B, and C are output. If the SINR is greater than the set threshold, the two nodes are allocated channels. At this time, after the data is transmitted according to the allocated channel, according to the data priority method in algorithm 2 (assuming that C data priority is greater than B), node E, which is within the communication radius of node C, receives the information of node C and sets its parent. The node position is C, and the iterative loop assigns transmission channels to the parent node in turn. In this paper, the allocation of time slots is performed after the channel is allocated to the parent node. In the time gap, it is necessary to find out all the nodes that are communicating in advance, and then check whether there is a parent node for data communication among the nodes that are currently communicating, and find out the communication child nodes according to the parent node of the communication to realize the transmission of the allocated channel of the node.
34.4 Simulation Experiment Delay perception performance of multi-channel scheduling algorithm: Demonstrate the effect of multi-channel scheduling, the effect of priority-based delay-aware scheduling algorithm on data transmission delay is mainly studied. It can be seen from experiment that in multi-channel scheduling, all types of data packet transmission delays have been significantly improved. In order to further reduce the data transmission delay, this experiment combines algorithm to discuss the impact of the priority-based delay-aware scheduling algorithm on data transmission delay. It can be seen from Fig. 34.4b that under single-channel and multi-channel transmission with high traffic load, the transmission delay performance with delayaware scheduling is significantly better than that without delay-aware scheduling. Comparing Fig. 34.4a, b, it can be seen that compared to the MHST algorithm, the
(a) Low load latency awareness (b) High load delay perception Fig. 34.4 Performance comparison of delay-aware scheduling under different loads (LB-MHST)
34 Multi-channel Scheduling for Intelligent …
291
LB-MHST algorithm improves the delay performance by 12% in single-channel scheduling, the delay performance increases by 15.2% at 8 channels, and at 16 channels. The delay performance is improved by 18%.
34.5 Conclusion Field measurement shows that low-power wireless communication in the application of infrastructure for intelligent information and intelligent collection will suffer from noise, interference, poor connectivity and fading due to the harsh and complex power grid environment. In addition, the influence of the proposed algorithm on the realtime communication of the grid master station is also analyzed. A large number of simulation experiments shows that the proposed algorithm can effectively reduce the communication delay in the smart grid environment. In future, we will make an indepth investigation on the multi-channel scheduling of WSN in the worse smart grid environment. And study how to combine the scheduling strategy proposed in this paper with the weighted fair scheduling scheme to meet the fairness requirements in different intelligent information intelligent collection application scenarios. Acknowledgements This work was financially supported by the science and technology project to State Grid Corporation “Research on the key technologies of power grid disaster intelligent perception and emergency command (5700-202019185A-0-0-00)”.
References 1. Liu, Z.Y.: Technology of Smart Grid, pp. 56–68. China Electric Power Press, Beijing (2010) 2. Wang, Y.M.: Research framework of robust smart grid technology standard system. Autom. Electr. Syst. 34, 1–6 (2010) 3. Li, W., Ding, J., Yao, J.G.: Discussion in development of power system automation. Autom. Electr. Syst. 34, 24–27 (2010) 4. Yaghmaee, M.H., Yousefi, Z., Zabihi, M. et al.: Quality of service guarantee in smart grid infrastructure communication using traffic classification. In: International Conference and Exhibition on Electricity Distribution, pp. 1–4. IET Press, Stockholm, Sweden (2013) 5. Xu, B.Y., Li, T.Y., Xue, Y.D.: Intelligent distribution network and distribution automation. Autom. Electr. Syst. 33, 38–41 (2009) 6. Bari, A., Jiang, J., Saad, W. et al.: Challenges in the smart grid applications: an overview. Int. J. Distrib. Sensor Netw. 2014, 1–11 (2014) 7. Musham, R., Kala, S.M., Muthyap, P. et al.: Near optimal channel assignment for interference mitigation in wireless mesh networks. In: 2016 IEEE International Conference on Advanced Networks and Telecommunications Systems, pp. 1–6. ANTS Press, Bangalore, India (2016) 8. Suryawanshi, R.: H-WSN with maximized QoS using secure data aggregation. In:2017 2nd International Conference on Contemporary Computing and Informatics, pp. 149–154. IC31 Press, Noida, India (2017)
292
X. Bao et al.
9. Jayalakshmi, G., Nitya, N., Bsanakar, R.M.: Congestion based routing metric and channel assignment scheme for WMN. In: Proceeding of Second International Conference on Electrical, Computer and Communication Technologies, pp. 1–6. ICECCT Press, Tamil Nadu, India (2017) 10. Ahmed, A.A., Fisal, N.: A real-time routing protocol with load distribution in wireless sensor networks. Comput. Commun. 31, 3190–3203 (2008) 11. Incel, O.D., Ghosh, A., Krishnamachari, B. et al.: Fast data collection in tree-based wireless sensor networks. IEEE Trans. Mob. Comput. 11, 86–99 (2011) 12. Han, B., Han, T., Chen, L.J. et al.: Design and implementation of network model checking method for master station of power distribution network. Autom. Electr. Power Syst. 40, 75–79 (2016)
Chapter 35
Anti-interference Performance of Transmit Diversity Techniques for Satellite System Ya Wen and Xin Wu
Abstract In order to investigate the anti-interference performance of SORTD (space orthogonal-resource transmit diversity) for PUCCH (Physical Uplink Control Channel) format 3 in satellite system, this paper gives overhead analysis of transmit diversity for PUCCH format 3. The result of simulation shows that there is no strong requirement for enhancement of transmit diversity for PUCCH format 3 from the overhead point of view. Thus, a new scheme of transmit diversity for PUCCH format 3 should not be introduced in satellite system unless it brings little complexity on specification and implementation.
35.1 Introduction In 3GPP Release 10, SORTD is supported as the transmit diversity scheme for PUCCH format 3 in satellite system. The amount of required PUCCH format 3 resources for SORTD is doubled compared to single antenna port transmission, and thus there was some concern on the PUCCH format 3 overhead with SORTD. One of the objectives in carrier aggregation work items is investigating the possible improvements including the anti-interference performance of transmit diversity scheme for PUCCH format 3 in satellite system [1]. In the RAN1 #66bis meeting, anti-interference performance of transmit diversity for PUCCH format 3 in satellite system was discussed. It was concluded that additional analysis on the overhead needed to be performed to identify whether enhancement of transmit diversity for PUCCH format 3 in satellite system is needed [2]. The contribution firstly discusses the method of performing the overhead analysis and then provides the overhead analysis of the anti-interference performance for PUCCH format 3 in satellite system.
Y. Wen (B) · X. Wu Chongqing College of Electronic Engineering (CQCEE), Chongqing 401331, China © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_35
293
294
Y. Wen and X. Wu
The rest of this paper is organized as follows. In Sect. 35.2, we discuss the method of performing the overhead analysis of PUCCH format 3 in satellite system. Then the overhead analysis of transmit diversity for PUCCH format 3 in satellite system is described in detail in Sect. 35.3. Simulation results are presented and discussed in Sect. 35.4. This paper is concluded in Sect. 35.5 at last.
35.2 Discussion on Overhead Analysis Method For the method of performing overhead analysis, the first question to be resolved is whether the PUCCH format 3 in satellite system overhead indicates the resources reserved semi-statically or the resources used in a given subframe. If all the unused PUCCH format 3 in satellite system resources can be used for other purposes like PUSCH (Physical Uplink Shared Channel) transmission in a given subframe, antiinterference performance should be evaluated based on the PUCCH format 3 in satellite system resources used in a given subframe [3]. However, usually, it is difficult to guarantee that all the unused PUCCH format 3 in satellite system resources can be used for PUSCH transmission because it would restrict the scheduling a lot. Thus, it is reasonable to evaluate the overhead based on the PUCCH format 3 in satellite system resources reserved semi-statically. In Rel-10, the typical application scenarios of PUCCH format 3 in satellite system mainly include the cases with more than two configured serving cells. In time division dual (TDD), one or two configured serving cells with heavy DL configurations can also be the case. Whether PUCCH format 3 in satellite system will be configured to a user equipment (UE) mainly depends on the traffic requirement and the PUCCH channel condition. For traffic requirement, it is difficult to evaluate how frequent the typical application scenarios will happen because it depends on the traffic model. In addition, even though the typical application scenarios need to be configured because of the traffic requirement, the scenario or PUCCH format 3 in satellite system cannot be configured unless the performance target can be met under the PUCCH channel condition. Thus, for simplicity, we can evaluate the anti-interference performance of PUCCH format 3 in satellite system only from the PUCCH channel condition point of view. From this point of view, the following two aspects need to be evaluated first: 1.
2.
PUCCH geometry The PUCCH geometry determines the number of UEs that can be configured with PUCCH format 3 in satellite system. A system level simulation should be used to get the PUCCH SINR (signal to interference plus noise ratio) geometry. When performing the simulation, we should consider interference and UL PUCCH power control. Link-level performance of PUCCH format 3 in satellite system Based on the PUCCH geometry and the link-level performance of PUCCH format 3 in satellite system with single antenna port (SAP) transmission and SORTD, we can get the percentage of UEs that can be configured with SAP
35 Anti-interference Performance of Transmit Diversity …
295
and the percentage of UEs that can be configured with SORTD, then we can get the PUCCH format 3 in satellite system resources that need to be reserved semi-statically for SORTD, assuming all the UEs with high enough geometry is configured with PUCCH format 3 in satellite system. Based on the PUCCH format 3 in satellite system resources reserved semi-statically for SORTD [4], we can analyze whether enhancement of transmit diversity for PUCCH format 3 in satellite system is needed or not.
35.3 Overhead Analysis of Transmit Diversity for PUCCH Format 3 in Satellite System In this section, we evaluate the overhead of PUCCH format 3 in satellite system based on the method discussed in Sect. 35.2. PUCCH geometry and link-level performance of PUCCH format 3 in satellite system are evaluated firstly. Interference and PUCCH power control in our simulation are set as follows: 1.
2.
For interference The cases with maximum 1–5 UEs multiplexed on the same physical resource block (PRB) are simulated. In ideal condition, the inter-cell interference mainly comes from UEs using the same orthogonal cover code (OCC) as the desired UE. However, UEs using different OCCs also can be interferers considering some factors like the uplink (UL) reception window. In addition, the interference randomization for PUCCH format 3 in satellite system also has impact on the actual interference. For simplicity, we can evaluate the cases with different number of multiplexed UEs on the same PRB. The multiplexed UEs on the same PRB in other cells are the interferers. More multiplexed UEs can mean heavier interference to the desired UE. For PUCCH power control The setting of the UE transmission power PPUCCH in sub-frame i is as follows: PPUCCH (i) = min PCMAX P0_PUCCH + P L + h n CQI , n HARQ , n SR + F_PUCCH (F) + TxD F + g(i)
(35.1)
where PCMAX is maximum transmitting power of UEs, P0_PUCCH is Power reference value, P L is the biggest loss of down channel measured, h n CQI , n HARQ , n SR is the transmitting power offset of channel quality indicator (CQI), hybrid automatic repeat request (HARQ) and scheduling request (SR), TXD is the power offset of modulation encoding mode and data type, g(i) is the adjustment value of closed-loop power control. Then
296
Y. Wen and X. Wu
1 P0_PUCCH = P0_PUCCH +h n CQI , n HARQ , n SR + F_PUCCH (F) + TxD F + g(i) (35.2) 1 And P0_PUCCH is determined according to the following steps: Get the target PUCCH received SINR (T-SINR). In our simulation, the target PUCCH received SINR is 3 dB. 1 , with which the most UEs could fulfil the target PUCCH Search the best P0_PUCCH 1 could ensure more UEs have a reliable received SINR. That is, the best P0_PUCCH PUCCH condition.
35.4 Simulations The simulation results lead to the following observation: More multiplexed UEs on the same PRB, which means the inference is heavier, lead to worse PUCCH SINR CDF. Figure 35.1 shows the PUCCH SINR cumulative distribution function (CDF) for the cases with different numbers of multiplexed UEs on the same PRB. The link-level performance of PUCCH format 3 in satellite system with SAP and SORTD was evaluated in several contributions [5, 6]. Based on the simulation results in [5, 6], we can get the temporary minimum requirements for some typical
Fig. 35.1 CDF for the cases with different number of multiplexed UEs on the same PRB
35 Anti-interference Performance of Transmit Diversity …
297
Table 35.1 The minimum requirements (SNR [dB]) for different A/N payload sizes 6 bits
10 bits
16 bits
20 bits
Single antenna port (SAP)
–5.1
–2.27
0.93
2.9
SORTD
–6.91
–5.1
–2.39
–1.39
Table 35.2 Percentage of UE that can support PUCCH format 3 with SAP and SORTD 6 bits (%) 10 bits (%) 16 bits (%) 20 bits (%) Percentage of UEs that can support PUCCH 61–88 format 3 with SAP
49–80
32–70
19–60
Percentage of UEs that can support PUCCH 67–91 format 3 with SORTD
61–88
50–81
46–79
A/N payload sizes as shown in Table 35.1. Here the minimum requirement for each A/N payload size is the maximum required SINR among the simulation results in [5, 6]. Based on the simulation results for PUCCH SINR geometry and the antiinterference performance of PUCCH format 3 in satellite system, we can get the percentage of the UEs that support PUCCH format 3 in satellite system with SAP and the percentage of UEs that can support PUCCH format 3 in satellite system with SORTD as shown in Table 35.2. If the anti-interference performance of PUCCH format 3 in satellite system with SAP can meet the target SINR requirement, it is reasonable to assume that SORTD will not be configured, because configuring SORTD requires doubled resource. Then we can further get the percentage of UEs that will be configured with PUCCH format 3 in satellite system with SAP and the percentage of UEs that will be configured with PUCCH format 3 in satellite system with SORTD as shown in Table 35.3. According to Table 35.3, we can see that the percentage of the UEs that will be configured with PUCCH format 3 in satellite system with SORTD is small when the A/N payload size is 6 bits, thus the increase of the PUCCH format 3 in satellite system overhead brought by SORTD is marginal. When the A/N payload size is more than 6 bits, the percentage of the UEs that will be configured with SORTD is approximate 8% to 27%, which is large and thus it may have some impact on the PUCCH format 3 in satellite system overhead. Table 35.3 Percentage of UE that will be configured with PUCCH format 3 with SAP and SORTD 6 bits (%)
10 bits (%)
16 bits (%)
20 bits (%)
Percentage of UEs that will be configured with PUCCH format 3 with SAP
61–88
49–80
32–70
19–60
Percentage of UEs that will be configured with PUCCH format 3 with SORTD
3–6
8–12
11–18
19–27
298
Y. Wen and X. Wu
35.5 Conclusions In this contribution, we discuss the method of performing the anti-interference performance of PUCCH format 3 in satellite system firstly and then provide the overhead analysis of transmit diversity for PUCCH format 3 in satellite system. Based on the evaluations and analysis, we can see that there is no strong requirement for enhancement of transmit diversity for PUCCH format 3 in satellite system from the overhead point of view. Thus, a new scheme of transmit diversity for PUCCH format 3 in satellite system should not be introduced unless it brings little complexity on specification and implementation. Acknowledgements This work is supported by Research and Application of High Integrated Modular Avionic Systems (cstc2019jscx-zdztzxX0045).
References 1. RP-110451: LTE Carrier Aggregation Enhancements, Nokia Corporation, Nokia Siemens Networks, Kansas City, USA, March 15th–18th (2011) 2. Iordache, A.V., Marghescu, I.: Transmit diversity in LTE network. Telecommunications Forum (TELFOR), 2013 21st. IEEE, pp. 145–148 (2013) 3. 3GPP.: Draft Report of TSG RAN WG1 #66bis v0.1.0, Zhuhai, China, 10th–14th October (2011) 4. 3GPP TS 36.213. Physical layer procedures 5. R1-113212: Transmit Diversity for PUCCH format 3, Huawei, HiSilicon, Zhuhai, China, October 10–14 (2011) 6. R1-113553: Transmit diversity for PUCCH Format 3, CATT, Zhuhai, China, October 10–14 (2011)
Chapter 36
Wireless Sensor Energy Harvesting and Management Huanan Zhang and Feng Wang
Abstract This paper analyzes the energy harvesting and management of wireless sensor network nodes and studies the principle, technology and method of energy collection of wireless sensor network nodes. Energy management of wireless sensor network nodes should be solved from two aspects of energy saving and energy supply. The harvesting principles and methods of various energy sources in the environment, including solar energy, wind energy, sound energy, vibration energy, thermoelectricity and electromagnetic field energy, are analyzed. Sensor nodes should absorb energy from the environment in as many ways as possible to ensure long-term, stable and reliable operation of sensor nodes.
36.1 Introduction Wireless sensor network (WSN) is self-organizing, miniaturized and aware of the outside world. It integrates many new technologies, such as sensors, embedded systems, communication and power supply. Wireless sensor network nodes are usually composed of sensors and communication circuits and can be placed in factories, farmland or battlefield and other dangerous or inconvenient places to complete positioning, measurement, control and other functions. Can at any time, any place through data collection, processing, analysis, dissemination to provide a new information channel, so that people get more detailed and reliable information, in industrial control, traffic management, environmental monitoring, space exploration and other fields has a potential and extensive application needs [1]. Wireless sensor network node is stationary and, possibly in the wild environment, is not allowed to change the battery; therefore, wireless sensor network node is the energy management problem for prolonging the lifetime of wireless sensor H. Zhang School of Data and Computer Science, Guangdong Peizheng College, Guangzhou 510830, China F. Wang (B) School of Computer Engineering, Hubei University of Arts and Science, XiangyangHubei Province 441053, China e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_36
299
300
H. Zhang and F. Wang
network application and reduce the cost of the key, to become one of the core problems in the research of wireless sensor network, involves two aspects, namely energy consumption and energy problems. Therefore, in order to solve the problem of energy management of wireless sensor network nodes, it is necessary to conduct in-depth and detailed research from these two aspects. At present, there are many researches on solving the problem of energy consumption. For example, in order to effectively utilize the existing energy resources and extend the life cycle of the network, various optimized routing communication protocols are studied. It is impossible for wireless sensor network nodes to work normally for a long time only by various optimization and consumption reduction methods. When the energy consumption has been reduced to a certain limit by various measures, no matter how hard people try, they will not get a better effect. Therefore, we must study from the perspective of energy supply and take effective methods to provide continuous energy supply for wireless network sensors. Wireless sensor network nodes can also obtain and store energy from their environment, so research on how to effectively collect and store energy from the environment and energy collection methods have attracted more and more researchers’ attention [2].
36.2 Requirements for Energy Harvesting Wireless sensor networks generally consist of a large number of sensor nodes scattered over a certain area, usually powered by batteries. However, due to the limitations of node size, the configured battery can provide very limited energy. At the same time, because the sensor nodes are often in harsh environments or inaccessible environments, and the number of sensor nodes is very large, it is not possible to replace the battery for each node. In order to prolong the life of sensor nodes, it is necessary to take various methods to replenish the energy of sensor nodes. The sensor nodes harvest energy from the environment, known as energy harvesting. Wireless sensor nodes collect all available energy from the environment and store it. When the node needs energy, the energy is taken out of the storage unit and transformed to get the energy needed on the node [3]. According to the environment of sensor nodes, the energy collected in the environment is also different, so the energy collection method of a single energy is difficult to ensure that all nodes in the wireless sensor network can reliably obtain the required energy. Therefore, it is necessary to set up two or more energy collection methods for each sensor node, which requires that a comprehensive energy collection power supply should be configured as far as possible in a wireless sensor network node with limited space according to the possible energy types in the working environment of the node [4].
36 Wireless Sensor Energy Harvesting and Management
301
36.3 Analysis of Energy Harvesting Technology There may be a variety of potential energy available in the physical space we live in, such as solar (light) energy, wind energy, heat energy, mechanical vibration energy, sound energy, electromagnetic field energy and so on. How to collect and store the energy in small sensor network nodes is one of the focus problems that many scientists have studied hard in recent years, and some progress has been made at present. Among them, the energy harvesting technology using mechanical vibration and light energy is more studied, and the products of related devices appear, which has a good application prospect.
36.3.1 Solar (Light) Energy Collection New advances in optoelectronic materials make light energy collection, a new method of power source for wireless network sensors. The installation and operation cost of optoelectronic components can also be greatly reduced with large-scale applications. The basic principle of photoelectric acquisition is to use photoelectric materials to absorb a large number of photons. If enough photons can activate the electrons in the photocell, electrons can be obtained after proper structural design. Photoelectric components act as decoders, generating voltages under light irradiation, which, combined with the corresponding adjustment and storage circuits, can supply power to the load. The amount of electric quantity is a function of the collected light energy. To obtain more electric quantity, the photoelectric element is usually placed in a welllit environment and the area of illumination is increased. Usually photovoltaic cells can produce voltage DC 0.5 V, but the actual voltage output varies with the operating temperature. Generally speaking, the lower the temperature, the higher the output voltage, the stronger the light, the greater the current output. Optoelectronic technology has evolved from the initial manufacture of silicon crystals to today’s deposition of tiny particles on photosensitive substrates. This new material can be used indoors or outdoors; it is easy to install in weight, and is less affected by ambient temperature, making it ideal for powering small, remote sensors. Solar collection power sources using super capacitors as energy buffers are a universal and representative model that is well suited for a wide range of sensor network applications, as shown in Fig. 36.1.
36.3.2 Wind Energy Harvesting The wind energy collection device consists of a rotating shaft and a number of blades. One end of the blade is fixed on the rotating axis and radially extends to all sides along the rotating axis. The blades form an oblique Angle to the shaft. When the wind
302
H. Zhang and F. Wang
Fig. 36.1 Energy harvesting sensor node
blows head-on against the blade, the blade rotates on its axis under the action of the wind. The shaft rotates, collecting energy from the wind. The wind energy collecting effect is related to the wind side. The effect of wind from different directions is very different. The wind force also has a big impact on wind harvesting.
36.3.3 Harvesting of Thermoelectric Energy Research on thermoelectric technology began in the 1940s and reached its peak in the 1960s with the successful realization of long-term power generation on spacecraft. The thermoelectric generator has the characteristics of small size, light weight, no vibration, no noise, reliable performance, less maintenance and can work for a long time in extreme harsh environment. It is suitable for small power supply of less than 5 W, for a variety of unsupervised sensors, small short-range communication devices and medical and physiological research instruments. At present, the relevant products have entered the practical stage. In recent years, thermoelectric generator has also shown a good application prospect in civil use. Through the in-depth study of thermoelectric conversion materials and the development of new materials, it has become the core content of thermoelectric technology research to continuously improve the thermoelectric performance and improve the electric output power under the condition of constant heat source. Scientists have invented a new type of battery that uses human temperature differences to generate electricity, which could provide longterm power for portable miniature electronic instruments, eliminating the trouble of charging or replacing batteries. As long as there is a temperature difference of 5 °C between the human skin and clothes, you can use this battery to provide enough energy for an ordinary watch.
36 Wireless Sensor Energy Harvesting and Management
303
36.3.4 Collection of Sound Energy Sound energy is a form of energy. Its essence is the transfer and transformation of mechanical energy through the media and in the form of wave after the object vibrates. Conversely, the transfer and transformation of other energies can also be reduced to mechanical energy to produce sound. The change can be reversed. To convert sound energy into electricity when sound waves encounter obstacles, it is possible to design and build resonators that collect sound when the input noise energy they collect is converted into electricity storage. Conventional engines are limited by thermodynamics and engine complexity. The most efficient engines are typically giant turbines used in power stations. Small sound-energy engines are up to 10% more powerful than the largest, most efficient turbines and have no moving parts and no maintenance.
36.3.5 Magnetic Energy Harvesting In the magnetostatic case, electric current and magnetic field always exist together. Thus,Therefore, the magnetic energy can be regarded as related to electric currents and stored in the magnetic field. However, scientific practice has proved that magnetic field is a special form of matter, which can exist without electric current. The changing electric field can also produce a magnetic field, which also has energy and its field energy density is the same as that of magnetostatic. In general, changing electromagnetic fields propagate in the form of waves, accompanied by energy transfer. The electromagnetic energy recovery system is mainly composed of receiving antenna, rectifier circuit and energy management module, among which the performance of antenna and rectifier circuit is the key factor affecting the performance of the whole system. In this system, the receiving antenna determines how much electromagnetic energy can be collected, and the rectifying circuit determines how much direct current energy can be converted into electromagnetic energy that can be directly used. On the one hand, the electromagnetic energy distributed in the environment has location, frequency uncertainty and power inhomogeneity, and the current design of electromagnetic energy recovery system mainly uses single-frequency or multifrequency high-gain directional antenna to solve the recovery problem. However, this scheme often requires high alignment of the receiving and receiving antennas, and the actual relative position between the energy source and the receiving antenna is often uncertain. On the other hand, there are two main problems in the existing rectifier circuit schemes: impedance matching, secondary rectification and impedance compression are adopted to improve the conversion efficiency, but the input power is too high when the optimal rectification efficiency is achieved; The nonlinear action of diodes makes the input power range of the rectifier circuit narrow when efficient rectification. Therefore, for the practical application scenarios with uncertain
304
H. Zhang and F. Wang
electromagnetic energy, the existing recovery schemes still fail to solve the current problems well [5].
36.3.6 Comprehensive Harvesting of Energy from Various Sources We live in an environment in which there are many forms of energy. Energy takes different forms in different environments. In order for each sensor node to obtain as much energy as possible from its environment, it is necessary to design such an energy collection system, which cannot collect energy from only one kind of energy, otherwise, the node will not work reliably for a long time once the environment is short of this kind of energy. Therefore, it is necessary to integrate multiple energy harvesting methods on each node. Of course, the difficulties are not difficult to imagine, mainly manifested in the following aspects: (1) various energy recovery technologies and methods are not mature at present, and researchers need to carry out a lot of innovative research; (2) Various energy harvesting components must meet the stringent size requirements of sensor network nodes; (3) Ensure that the various energy harvesting components work in concert and that the collected energy is effectively stored.
36.4 Energy Management The energy collected must be managed effectively so that it can be continuously and effectively supplied to the sensor, otherwise the energy collected will be useless. Power management is a hot topic in energy utilization. Dynamic power management is a potential method to save power. The storage element of power supply is battery or capacitor. Chemical battery has become the first choice in network sensor because of its stable and effective storage. The limitation of collecting energy into the battery is the battery life problem. How to make the battery life meet the user’s acceptable expectation and adopt the correct charging method to extend the battery life is the key problem. Dynamic power management and dynamic voltage dispatching are adopted in power management of networked sensors, which is an effective design method to reduce the energy consumption of the system. Dynamic power management (DPM) is a design method to reduce power consumption by dynamically allocating system resources and completing system tasks with the least components or the least workload. Dynamic power management technology includes a series of methods to enable the system to achieve efficient energy saving. These methods control the “power management” of whether system components enter a low-power state when they
36 Wireless Sensor Energy Harvesting and Management
305
are idle. The basic premise of the application of dynamic power management technology is that the system components have different workload during the working time. This is true of most systems. Another premise is that fluctuations in the workload of systems and components can be predicted with some degree of confidence. In this way, it is possible to switch between states of energy consumption, and the system cannot consume excessive energy within the observed and predicted time of the workload. In the network sensor power management, its working characteristics are very suitable for dynamic power management. Because the nodes of the sensor network are not working all the time, they spend a lot of time in sleep mode, which requires low power consumption. When the sensor collects the data, the controller processes the data and carries on the data communication, the system power consumption increases. When entering idle, the device can be shut down and enter the low-energy sleep state; when the request is received again, the device is evoked [6]. In a dynamic power management system, the working states of different components must dynamically adapt to different performance requirements. Only in this way can the energy wasted in idle time or in useless components be minimized. For the judgment of power management time, it is necessary to use a variety of prediction methods, such as static prediction method and dynamic prediction method. It is necessary to predict the coming workload according to the historical workload and decide whether to change the working state. The main principle of dynamic voltage scheduling is to adjust the supply voltage dynamically based on the load state to reduce the power consumption of the system, which has been applied in portable personal mobile devices. Sensor nodes on the embedded operating system is responsible for scheduling request to receive services from different task queue, and realtime monitor the processor utilization and task queue length, the load observer based on the sequence of these two parameters value computation load of nominal value, DC/DC converter with reference to the value of the output amplitude of voltage, support the work of the processor. This actually constitutes a closed-loop feedback system, and the method of control theory can be used to design each module.
36.5 Conclusion Wireless sensor network is widely used. In order to prolong the life cycle of sensor nodes, improving the energy efficiency of sensor nodes has always been an important research topic. In this paper, the energy collection and management of wireless sensor network nodes are analyzed and discussed. Solar energy, wind energy, sound energy, magnetic energy and so on are introduced and analyzed. In order to make the sensor nodes work stably and reliably for a long time, the most effective method is to adopt various methods to obtain energy from the environment and reduce energy consumption and improve energy efficiency through energy management.
306
H. Zhang and F. Wang
References 1. Wang, Q., Dai, H. N., Li, X., Wang, H., Xiao, H.: On modeling eavesdropping attacks in underwater acoustic sensor networks. Sensors 16(5), 721 (2016) 2. Wang, Z., Han, G., Qin, H, Zhang, S., Sui, Y.: An energy-aware and void-avoidable routing protocol for underwater sensor networks. IEEE Access, 8(6), 7792–7801 (2018) 3. Gherbi, C., Aliouat, Z., Benmohammed, M.: An adaptive clustering approach to dynamic load balancing and energy efficiency in wireless sensor networks. Energy, 114, 647–662 (2016) 4. Long, J., Dong, M. X., Ot, K., Liu, A. F.: Green TDMA scheduling algorithm for prolonging lifetime in wireless sensor networks. IEEE Syst. J. 1–10 (2015) 5. Mohanty, P., Kabat, M.R.: Energy efficient reliable multi-path data transmission in WSN for healthcare application. Int. J. Wirel. Inf. Netw. 23, 162–172 (2016) 6. Lal, C., Petroccia, R., Pelekanakis, K., Conti, M., Alves, J.: Toward the development of secure underwater acoustic networks. IEEE J. Ocean. Eng. 42, 1075–1087 (2017)
Chapter 37
Stacked Sparse Autoencoder-Based WLAN Fingerprinting Localization Junhang Bai, Yongliang Sun, and Quan Zhang
Abstract Currently, deep learning algorithms have been used in a variety of fields because their performance usually outperforms traditional machine learning algorithms. In this study, we present a stacked sparse autoencoder (SSAE)-based wireless local area network (WLAN) fingerprinting localization algorithm. The hidden layer dimension of a basic autoencoder (AE) is limited by its input dimension, which could affect its performance. By contrast, the hidden layer dimension of a sparse autoencoder can be set to be greater than its input dimension, so we train two sparse autoencoders in turn for unsupervised received signal strength indicator (RSSI) feature learning. Then we stack the encoders and also add a linear regression layer as an output layer to the SSAE structure. We obtain the SSAE fingerprinting localization algorithm after supervised fine-tuning. We evaluate our presented SSAE localization algorithm in an actual indoor office scenario. The localization results show that our presented SSAE fingerprinting algorithm achieves a better performance compared with some other popular fingerprinting localization algorithms.
37.1 Introduction With the development of mobile communication and proliferation of smartphone applications, location-based service (LBS) has been paid intensive attentions and shown great application potentials in the venues like shopping mall, hospital, and so on [1]. Due to the limited performance of satellite and cellular-based localization systems indoors, different localization systems for indoor scenarios have been implemented using ultrasound, Bluetooth, ultra-wideband (UWB), and wireless local area network (WLAN) [2]. Among these existing systems, WLAN fingerprinting localization has been favored because of the widely available WLAN access points (APs) and satisfactory localization performance [3]. At present, various fingerprinting localization algorithms have been utilized. For instance, weighted k-nearest neighbors (WKNN) and multi-layer perceptron (MLP) [4]. Nowadays, deep learning algorithms J. Bai · Y. Sun (B) · Q. Zhang School of Computer Science and Technology, Nanjing Tech University, Nanjing 211816, China e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_37
307
308
J. Bai et al.
have demonstrated its strong power in massive data processing, feature extraction, and so on [5] and has been applied in the field of WLAN fingerprinting localization. [6, 7]. Thus, in this study, we present a stacked sparse autoencoder (SSAE)-based WLAN fingerprinting localization algorithm.
37.2 System Model 37.2.1 System Overview Just like the basic fingerprinting localization system, the SSAE-based WLAN fingerprinting localization system also can be divided into two steps that are the offline and online step. Firstly, in the offline step, reference points (RPs) are selected and their locations are recorded. At these RPs, received signal strength indicator (RSSI) are recorded to establish a fingerprint database called radio-map. Specifically, the SSAE fingerprinting localization algorithm is trained by employing the radio-map. The SSAE structure can be constructed by stacking the encoders of sparse autoencoders after the unsupervised feature learning, and then its weights and biases are optimized through the supervised fine-tuning. In the online step, the measured RSSI vector can be inputted into the trained SSAE to calculate the user’s location vector.
37.2.2 Fingerprinting Localization Using SSAE Unsupervised Feature Learning. A basic AE is a kind of neural network that has one hidden layer. It consists of an encoder E(r ) and a decoder D(r ). Through training, the decoder output vector is expected to approach to the encoder input vector, that is D[E(r )] ≈ r . The encoder output vector a and the decoder output vector rˆ are computed as follows:
a = E(wr + b) rˆ = D wT a + b
(37.1)
where r is the input vector of the AE, w is the weight matrix between the input layer and the hidden layer, wT is the transposition of w and is the weight matrix between the hidden layer and the output layer, b is the bias vector of the hidden layer, b is the bias vector of the output layer. We employ the back propagation (BP) to update weight matrices and bias vectors of the AE to minimize the loss function JAE between the r and rˆ . The optimized weight matrices and bias vectors of the AE are obtained when the loss function JAE is reduced. The JAE can be computed as follows:
37 Stacked Sparse Autoencoder-Based WLAN …
JAE =
309
N 1 r n − rˆ n 2 N n=1
(37.2)
where r n is the nth input vector, rˆ n is the nth output vector, N is the number of input vectors, and ·2 is 2-norm. An AE can be used for dimension reduction or feature extraction of RSSI data. In this progress, the dimension reduction of the hidden layer makes the network learn an RSSI compressed representation. If only a few APs can be detected by a terminal device at the same time, the dimension of input RSSI vector could be limited. Then the dimension reduction of the hidden layer of the AE will have a serious influence on the network performance. To solve this problem, we exploit a sparse autoencoder for unsupervised feature learning. The dimension of the hidden layer of the sparse autoencoder is not limited by the dimension of input RSSI vector, which will improve the performance of the supervised fine-tuning. A sparse autoencoder can be derived through adding a sparsity restriction to an AE’s hidden layer. Its hidden layer dimension can be set to be greater than its input dimension. The sparsity restriction can be considered that the hidden layer neurons are 0 in most cases. On this basis, we add a sparsity penalty factor JKL for restricting the sparsity to the loss function JAE . First, we compute the mean of the activation ρq of the qth neuron in the hidden layer of all the N input vectors of a sparse autoencoder. It can be calculated as follows: ρq =
N 1 h nq N n=1
(37.3)
where h nq is the qth hidden layer neuron output of the nth RSSI vector. Then we use a sparsity parameter ρ closed to 0 to approximately enforce the sparsity restriction ρq = ρ. Meanwhile, we also use relative entropy called Kullback– Leibler (KL) divergence as a regularization term, with which the sparsity penalty factor JKL is computed as follows: JKL =
Q q=1
1−ρ ρ ρ log + (1 − ρ) log ρq 1 − ρq
(37.4)
where Q is the number of neurons in the hidden layer. Besides, we also add a weight penalty factor JWP to the loss function to avoid over-fitting and the weight penalty factor JWP is calculated as follows: JWP =
Q L l=1 q=1
2 wlq
(37.5)
310
J. Bai et al.
where wlq is the weight between the lth input and the qth hidden layer neuron, L is the dimension of the input vector. Thus, the sparse autoencoder’s loss function JLF can be calculated as follows: JLF = JAE + k1 JKL +k2 JWP
(37.6)
where k1 is the weighted parameter of the sparsity penalty factor and k2 is the weighted parameter of the weight penalty factor. An SSAE can contain multiple encoders of sparse autoencoders. In this study, we establish an SSAE with sparse autoencoder 1 and sparse autoencoder 2. We firstly train sparse autoencoder 1 using the standardized RSSI data. After that, we also train sparse autoencoder 2 using the hidden layer neuron outputs of the sparse autoencoder 1. We also add a linear regression layer as the output layer to the trained SSAE structure. Figure 37.1 reveals the construction of the presented SSAE. Supervised Fine-Tuning. In order to fine-tune the SSAE, we exploit the RPs’ RSSI and location vectors as the SSAE inputs and outputs, respectively. The output vector y of the SSAE is calculated as follows: y = w(3) a(2) + b(3)
(37.7)
where a(2) is the vector in the hidden layer 2, w(3) is the weight matrix between the vector in the hidden layer 2 and the output vector, b(3) is the bias vector of the output layer. We also employ the BP to fine-tune weight matrices and bias vectors of the SSAE to minimize the loss function JSSAE . The loss function JSSAE is computed as follows: JSSAE =
N 1 L n − yn 2 N n=1
Fig. 37.1 The construction of the presented SSAE
(37.8)
37 Stacked Sparse Autoencoder-Based WLAN …
311
where L n is the real location of nth RSSI vector, and yn is the calculated location by our presented SSAE using nth RSSI vector.
37.3 Experimental Setup and Results 37.3.1 Experimental Setup We obtained the RSSI and location data in an actual office scenario with 20.4 m width × 51.6 m length. The plan is shown in Fig. 37.2. We mounted 7 TP link TLWR845N APs at 2.2 m height. In order to establish the radio-map, we chose 116 RPs totally in the rooms and corridor, and then we recorded 120 RSSI vectors at every RP. Similarly, we also chose 90 test points (TPs) in the scenario and recorded 60 RSSI vectors at every TP to test our presented algorithm. We utilized a Meizu M2 smartphone at 1.2 m height for RSSI measurement with a rate of 1 RSSI vector per second. In the unsupervised feature learning phase, we set up the sparsity parameter ρ = 0.05 of the two sparse autoencoders. We set up the weighted parameters k1 = 1 and k2 = 0.001. At the same time, we generated the weight matrices of all the sparse autoencoders using a Gaussian distribution with mean μ1 = 0 and variance σ12 = 0.01, and set each component of the bias vectors of all the sparse autoencoders equal to 0.1. Additionally, we set up the neuron numbers Q 1 = 100 and Q 2 = 200 in the hidden layer 1 and 2, respectively. Also, we set up the iterative epochs ie1 = 300 and learning rate lr1 = 0.01. In the supervised fine-tuning phase, we generated the output layer weight matrix of the SSAE using a Gaussian distribution with mean μ2 = 0 and variance σ22 = 1, and set each component of the bias vectors of the output layer equal to 0.1. To fine-tune the weights and biases of the SSAE, we also set up the iterative epochs ie2 = 300, dropout rate dr = 0.2, and learning rate lr2 = 0.005. Fig. 37.2 The experimental scenario plan
312 Table 37.1 Performance comparison of different fingerprinting algorithms
J. Bai et al. Algorithm
Mean error (m)
Cumulative probability (%) Within 2 m error Within 3 m error
WKNN
4.20
37.9
54.4
MLP
2.55
45.1
69.3
DNN
2.45
48.9
72.2
SAE
2.34
51.4
73.0
SSAE
2.23
56.7
75.1
37.3.2 Experimental Results We compare our presented SSAE fingerprinting algorithm with some other popular fingerprinting algorithms: WKNN, MLP, deep neural network (DNN), and stacked autoencoder (SAE). Regarding the WKNN, we select 7 nearest RPs for localization. We set up the hidden layer neuron number of the MLP Q MLP = 100. The DNN, SAE, and SSAE have the same parameters and network structure. A total of 5400 testing RSSI vectors are used for testing these algorithms, and the computed localization results are listed in Table 37.1. The mean testing errors of the WKNN, MLP, DNN, SAE, and presented SSAE are 4.20 m, 2.55 m, 2.45 m, 2.34 m, and 2.23 m, respectively. The presented SSAE has the minimum mean error. Meanwhile, the cumulative probabilities of different fingerprinting algorithms are shown in Fig. 37.3. The cumulative probabilities of the SSAE with a localization error of 2 m and 3 m can reach 56.7% and 75.1%, respectively. The SSAE outperforms the other fingerprinting algorithms.
Fig. 37.3 The cumulative probability curves of different fingerprinting algorithms
37 Stacked Sparse Autoencoder-Based WLAN …
313
37.4 Conclusion We present an SSAE-based WLAN fingerprinting localization algorithm. Because the hidden layer dimension of a sparse autoencoder is not limited by the dimension of input RSSI vector, we exploit the sparse autoencoder and train two sparse autoencoders for unsupervised feature learning. Then we stack the encoders of the two sparse autoencoders to construct an SSAE structure and add a linear regression layer to the SSAE structure for fingerprinting localization. We train the SSAE using BP algorithm for supervised fine-tuning. The localization results reveal that our presented SSAE fingerprinting algorithm has the minimum mean testing error of 2.23 m compared with some other popular fingerprinting algorithms. Acknowledgements This work was supported by the National Natural Science Foundation of China under Grant 61701223 and also by the Natural Science Foundation of Jiangsu Province under Grant BK20171023.
References 1. Guo, X., Ansari, N., Li, L., Duan, L.: A hybrid positioning system for location-based services: design and implementation. IEEE Commun. Mag. 58(5), 90–96 (2020) 2. Gu, Y., Lo, A., Niemegeers, I.: A survey of indoor positioning systems for wireless personal networks. IEEE Commun. Surv. Tutor. 11(1), 13–32 (2009) 3. Zafari, F., Gkelias, A., Leung, K.K.: A survey of indoor localization systems and technologies. IEEE Commun. Surv. Tutor. 21(3), 2568–2599 (2019) 4. Sun, Y., Meng, W., Li, C., Zhao, N., Zhao, K., Zhang, N.: Human localization using multi-source heterogeneous data in indoor environments. IEEE Access 5, 812–822 (2017) 5. Ma, X., Yao, T., Hu, M., Dong, Y., Liu, J.: A survey on deep learning empowered IoT applications. IEEE Access 7, 181721–181732 (2019) 6. Khatab, Z.E., Hajihoseini, A., Ghorashi, S.A.: A fingerprint method for indoor localization using autoencoder based deep extreme learning machine. IEEE Sensors Letters 2(1), 1–4 (2018) 7. Li, Y., Hu, X., Zhuang, Y., Gao, Z., Zhang, P., El-Sheimy, N.: Deep reinforcement learning (DRL): another perspective for unsupervised wireless localization. IEEE Internet Things J. 7(7), 6279–6287 (2020)
Chapter 38
A Gridless Joint DOA and Range Estimation Method for FMCW MIMO Radar Jingyu Cong, Tao Lu, Zhang Bin, Xiaoli Jing, and Xianpeng Wang
Abstract The traditional two-dimensional multiple signal classification (2DMUSIC) has low accuracy and high computational complexity which is not effective in the joint direction-of-arrival (DOA) and range estimation with frequencymodulated continuous wave multiple-input multiple-output (FMCW MIMO) radar. In this paper, a gridless algorithm for joint DOA and range estimation is proposed. Firstly, the atomic norm denoising method is used to reconstruct the low-noise covariance of the received signal. Next, the DOA of targets is calculated by the total least square estimation of signal parameters via rotational invariance techniques (TLSESPRIT) algorithm and then reconstruct the steering vector matrix. Finally, the range of targets is obtained by the least square method. At the same time, the proposed method ensures the automatic matching of DOA and range. The effectiveness of the algorithm is proved by experimental data of 2.4G band FMCW MIMO radar which is more excellent than 2D-MUSIC.
38.1 Introduction Frequency-modulated continuous wave (FMCW) radars has been successful in many applications [1]. FMCW multiple-input multiple-output (MIMO) radar has an equivalent virtual antenna array which is pairing transmitting and receiving elements as virtual element independently [2], which enlarges the array aperture and enhances the estimation performance. As a solution to estimate DOA and range estimation jointly, two-dimensional (2D) algorithms have been proposed [2–4]. In [2], 2D fast Fourier transformation (FFT) which is a fast algorithm is used to estimate direction-of-arrival (DOA) and range, unfortunately, limited by the Rayleigh criterion of FMCW MIMO system, J. Cong · Z. Bin · X. Jing · X. Wang (B) State Key Laboratory of Marine Resource Utilization in South China Sea and School of Information and Communication Engineering, Hainan University, Haikou 570228, China e-mail: [email protected] T. Lu Beijing Institute of Aerospace Engineering, Beijing 100076, China © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9_38
315
316
J. Cong et al.
and the DOA resolution and range resolution are not very desirable. In [3], a 2D multiple signal classification (2D-MUSIC) method for the DOA and range imaging was presented, which was able to obtain a good performance in experiments, but just like the algorithm in [4], a huge covariance matrix and 2D spectral peak search are needed that lead to high computational complexity. In [4], a CAPON-based 2D algorithm has been proposed; however, due to the matrix inversion, computational complexity is O(M 3 ), and this method is hard to meet the calculation speed requirements either. Thus, this paper proposed a gridless algorithm for joint DOA and range estimation. Firstly, by using atomic norm denoising method and total least square estimation of signal parameters via rotational invariance techniques (TLS-ESPRIT) algorithm to get DOA and reconstruct the steering vector matrix. Next, get the paired range of targets by the least square method. Finally, the effectiveness of the algorithm is proved by experimental data of 2.4G band FMCW MIMO radar which is more excellent than 2D-MUSIC.
38.2 Radar Model Assume that a FMCW MIMO radar with an equivalent virtual array, which contains M equally spaced elements. The signal model of single snapshot is as follows [1]: ⎡
⎤
⎡
4π f c x1 sin θ1 c
4π f c x1 sin θ2 c
e e x1 (n) 4π f x sin θ 4π f x sin θ j c 2c 1 j c 2c 2 ⎢ x2 (n) ⎥ ⎢ ⎢ e e ⎥ ⎢ ⎢ ⎢ . ⎥ =⎢ .. .. ⎣ .. ⎦ ⎣ . . 4π f c x M sin θ1 4π f c x M sin θ2 x M (n) c c ej ej ⎡ ⎤ w1 (n) ⎢ w2 (n) ⎥ ⎢ ⎥ +⎢ . ⎥ ⎣ .. ⎦ j
j
4π f c x1 sin θk c
··· e 4π f c x2 sin θk c ··· ej .. .. . . j
··· ej
4π f c x M sin θk c
⎤ 4π BTs R1 n 4π f R ⎤⎡ j + cc 1 V1 e Tc ⎥ 4π BTs R2 n 4π f R ⎥⎢ j + cc 2 ⎥ Tc ⎥⎢ V e ⎥ 2 ⎥⎢ ⎥ ⎥⎢ .. ⎢ ⎥ ⎦⎣ . ⎦
Vk e
j
4π BTs Rk n 4π f R + cc k Tc
(38.1)
w M (n)
where f c denotes the carrier frequency, T is the cycle time, c is the velocity of light, T s is the sampling time, xi = d × (i − 1) is the distance from the reference element, d denote the interspacings of the virtual elements, B is the bandwidth, θk , Rk are the DOA and range of the kth target, respectively, V k is the amplitude of the kth signal, and wi (n) is white noise. Combined with Eq. 38.1, the model of multisnapshot can be written as: X = AS + N ∈ C M×L
(38.2)
38 A Gridless Joint DOA and Range Estimation …
317
where the X denotes multisnapshot data matrix, the steering matrix is given by A, S is the matrix of K signals, N denotes the white noise matrix, and · ∈ Cx×y denotes · is a x × y matrix.
38.3 DOA Estimation 38.3.1 Atomic Norm Denoising According to [5], the optimization problem of atomic norm denoising method can be expressed in the form of minimization as follows: 1 min τ ZA + X − Z22 Z 2
(38.3)
where ·A denotes the atomic norm, Z represents complex matrix, and τ is the regularization coefficient. This optimization problem can be changed into the following SDP: min τ ( D D,Z,u 2
+ u1 ) + 21 X − Z22
D ZH ≥0 s.t. Z H(v)
(38.4)
where Ro = H(v) is a low-noise covariance matrix with the structure of Toeplitz, which is composed of vector v = [v 1 , v 2 , . . . , v M ]T ∈ C M , and D is a positive definite symmetric matrix.
38.3.2 TLS-ESPRIT DOA Estimation The signal subspace is formed by eigen decomposition of Ro , and the k largest eigenvalues are taken as the signal subspace, and the following eigenvalue decomposition is calculated after being divided into E x and E y two parts: E Hx E x + E Hy E y = UU H
(38.5)
Decompose U into a submatrix of k × k.
E=
E 11 E 12 E 21 E 22
(38.6)
318
J. Cong et al.
Calculate the eigenvalue λi , i = 1, 2, . . . , k of Ψ = −E 12 E −1 22 , then the estimated value of DOA is obtained by the following equation: θˆi = arcsin
c angle(λi ) , i = 1, 2, . . . , k 4π f c d
(38.7)
38.4 Range Estimation Using the DOA estimation results of the previous section, we can reconstruct the steering vector matrix of the system: A = a(θˆ1 ), a(θˆ2 ), . . . , a(θˆk )
(38.8)
T where a(θˆ1 ) = e j4π fc x1 sin θi / c , . . . , e j4π fc x M sin θi / c . Multiply Eq. 38.2 to the left by
+
+
A , owing to A A ≈ I K ×K , and the following equation is established:
+
Y ≈ S+ A N
(38.9)
Let Y 1 = Y T , which every column is a noisy signal matrix transposition. According to the signal matrix form in Eq. 38.1, it is not difficult to know that the phase difference of adjacent elements in each column is = 4π BTs Tc ; at the same time, each element contains the phase of the first element. Therefore, the phase matrix of real value can be obtained by normalization. Y 1 (:, i) = Im ln Y 1 (:, i) Y 1 (1, i) , i = 1, 2, . . . , k
(38.10)
where Im{·} denotes imaginary part operator, (:, i) denotes ith column, and (1, i) denotes the first element of ith column. Constructing least square matrix
Pr =
1 1 ··· 1 2 · · · L
(38.11)
The distance estimation corresponding to each DOA estimate can be obtained by multiplying generalized inverse of P r to each column of normalized Y 1 (:, 1). Rˆ i = P r+ Y 1 (:, i) (2) , i = 1, 2, . . . , k
(38.12)
38 A Gridless Joint DOA and Range Estimation …
319
where {·}(2) is the second element of the column vector.
38.5 Simulations and Experiments In this section, the experimental data are used to compare the performance between proposed algorithm with 2D-MUSIC [3]. The 2.4G FMCW radar parameters and two actual target parameters are shown in Tables 38.1 and 38.2. Figure 38.1 is the 2.4G FMCW MIMO radar system. Figure 38.2 shows the experiment scene. The number of snapshots is 400. As the results shown in Fig. 38.3, the proposed algorithm is better than a traditional 2D-MUSIC. Table 38.1 2.4G FMCW radar parameters Parameter
Value
Ts
2.5×10−5
B
80 Mhz
s
Table 38.2 Two actual target parameters
Parameter
Value
Parameter
Value
fc
2.4 Ghz
T
2.06×10−4 s
M
16
d
0.08 m
Parameter
Value
Parameter
Value (m)
θ1
−8.5°
R1
41.9
θ2
0.2°
R2
28.1
Fig. 38.1 2.4G FMCW MIMO radar system
320
J. Cong et al.
Fig. 38.2 Experiment scene
Fig. 38.3 Comparison of target location experiments
38.6 Conclusion In this work, we designed a gridless algorithm for joint DOA and range estimation, by applying atomic norm denoising method, TLS-ESPRIT algorithm, and least square method. Although the performance of the proposed method is verified on experimental data which is much better than 2D-MUSIC, it is limited to stationary narrowband far-field targets. Therefore, in the future work, some more practical targets will be considered, such as narrow-band far-field moving targets and wide-band moving targets. Acknowledgements This work is supported by Young Elite Scientists Sponsorship Program by CAST (No. 2018QNRC001), Key Research and Development Program of Hainan Province (No. ZDYF2019011), the Scientific Research Setup Fund of Hainan University (No. KYQD(ZR) 1731), the Program of Hainan Association for Science and Technology Plans to Youth R&D Innovation
38 A Gridless Joint DOA and Range Estimation …
321
(No. QCXM201706), the Scientific Research Setup Fund of Hainan University (No. KYQD(ZR) 1731), and the National Natural Science Foundation of China (No. 61701144, No. 61861015, No. 61961013).
References 1. Huang, L., Wang, X., Huang, M., Wan, L.: An implementation scheme of range and angular measurements for FMCW MIMO radar via sparse spectrum fitting. Electronics 3(9), 389 (2020) 2. Feger, R., Wagner, C., Schuster, S., Scheiblhofer, S., Jager, H., Stelzer, A.: A 77-GHz FMCW MIMO radar based on an SiGe singlechip transceiver. IEEE Trans. Microw. Theory Tech. 57(5), 1020–1035 (2009) 3. Belfiori, F., Rossum Van, W., Hoogeboom, P.: 2D-MUSIC technique applied to a coherent FMCW MIMO radar. In: IET International Conference on Radar Systems (Radar 2012), pp. 1–6. IEEE, UK (2012) 4. Hamidi, S., Nezhad-Ahmadi, M., Safavi-Naeini, S.: TDM based virtual FMCW MIMO radar imaging at 79 GHz. In: 2018 18th International Symposium on Antenna Technology and Applied Electromagnetics (ANTEM), pp. 1–2. IEEE, Waterloo (2018) 5. Li, Y., Chi, Y.: Off-the-grid line spectrum denoising and estimation with multiple measurement vectors. IEEE Trans. Signal Process. 5(64), 1257–1269 (2016)
Author Index
B Bai, Junhang, 307 Bao, Xingchuan, 285 Bin, Xing, 37 Bin, Yu, 37 Bin, Zhang, 315
C Cao, Yazhuang, 121 Chang, Liu, 75 Chao, Kang, 21 Chao, Ye, 37 Chen, Chaozhou, 223 Chen, Tingran, 1 Chen, Yan, 251 Chu, Yuqian, 87 Cong, Jingyu, 315
D Deng, Xi, 241 Dong, Guoliang, 223 Dong, Xuan, 105
Guicheng, Shen, 29 Guo, Zhi Qiang, 159 Guo, Zhiqi, 149
H Haoping, Zhang, 75 Huang, Xuewei, 63 Hu, Chaolang, 63 Huiping, Si, 75 Hui, Xin, 233 Hu, Ronghui, 267
J Jia, Hong, 97 Jiang, Huaizhen, 13 Jiang, Tingyan, 121 Jian, Sun, 111 Jie, Chen, 75 Jing, Xiaoli, 315 Junhui, Wu, 75
K Kaiyan, Lin, 75 F Fang, Quan, 45 Feng, Haotian, 241 Feng, Yong, 171, 181 Fu, Yunfa, 171, 181
G Gao, Ruiyang, 1 Gao, Shijing, 129
L Lei, Rujie, 199, 207 Liang, Yanzhi, 87 Li, Haiying, 13 Li, Jing Tao, 7 Li, Man, 63 Li, Ruijun, 137 Li, Shengqing, 241
© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 L. C. Jain et al. (eds.), Smart Communications, Intelligent Algorithms and Interactive Methods, Smart Innovation, Systems and Technologies 257, https://doi.org/10.1007/978-981-16-5164-9
323
324 Li, ShuQin, 129 Liu, Fujia, 223 Liu, Jingyi, 279 Liu, Yingji, 97 Liu, Yu, 53, 191 Li, Weiwen, 87 Li, Xiao Jiang, 7 Li, Yong, 233 Li, Zhen Feng, 7 Li, Zhenxiang, 149 Long, Ziping, 13 Luo, Kaijia, 149 Lu, Tao, 315 Lu, Zhang, 149 Lv, Ziliang, 1
M Ma, Ruiheng, 199
P Pan, Zhenghao, 129 Peng, Lin, 285 Pu, Lirong, 13
Q Qian, Qian, 171, 181 Qiu, Run, 7
S Shi-ming, Li, 21 Shi, Yaqing, 137 Song, Hu, 45 Sun, Bin, 13 Sun, Li, 233 Sun, Yongliang, 307
T Tang, Yiduo, 251, 259
W Wang, Feng, 299 Wang, Jiaqi, 63 Wang, Lei, 143 Wang, Peng, 199, 207 Wang, Shuming, 121 Wang, Xianpeng, 315 Wang, Yawen, 171 Wei, Xinlei, 97
Author Index Wenjia, Long, 215 Wen, Ya, 293 Wu, Xin, 293
X Xia, Fei, 45 Xia, Haiying, 97 Xiang, Min, 191, 199, 207 Xinhua, You, 215 Xuan, Xin, 121 Xu, Min, 285 Xu, Shuquan, 97 Xu, Xiaoxiao, 267 Xu, Zhong, 105
Y Yang, Bo, 233 Yang, Jian, 285 Yang, Jianhua, 285 Yang, Wang, 29 Yi-jie, Wu, 21 Yi, Zhang, 37 Yuan, Li, 105 Yuan, Yang, 21 Yu, Chao, 267 Yu, Dai, 215 Yue, Haiyun, 63 Yu, Kun, 199
Z Zhang, Hua, 13 Zhang, Huanan, 299 Zhang, Jieming, 87 Zhang, Peiyong, 233 Zhang, Quan, 307 Zhang, Shuying, 251, 259 Zhang, Xin, 1 Zhang, Yanan, 149 Zhao, Suaisuai, 267 Zhao, Xu, 53 Zhao, Yafei, 233 Zhao, Zhiliang, 13 Zhao, Zizhuo, 259 Zhou, Chao, 159 Zhou, Jiayi, 181 Zhou, Lin, 251, 259 Zhou, Run, 191 Zhou, Xiaona, 233 Zhou, Yi Ren, 159 Zhu, Zhisong, 53